molecules
Review
Artificial Neural Networks for Pyrolysis, Thermal Analysis, and
Thermokinetic Studies: The Status Quo
Nikita V. Muravyev 1, * , Giorgio Luciano 2 , Heitor Luiz Ornaghi, Jr. 3 , Roman Svoboda 4
and Sergey Vyazovkin 5, *






Citation: Muravyev, N.V.;
Luciano, G.; Ornaghi, H.L., Jr.;
Svoboda, R.; Vyazovkin, S. Artificial
Neural Networks for Pyrolysis,
Thermal Analysis, and Thermokinetic
Studies: The Status Quo. Molecules
2021, 26, 3727. https://doi.org/
10.3390/molecules26123727
Academic Editor: Teobald Kupka
Received: 30 May 2021
Accepted: 15 June 2021
Published: 18 June 2021
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2021 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Molecules 2021, 26, 3727. https://doi.org/10.3390/molecules26123727
1 N.N. Semenov Federal Research Center for Chemical Physics, Russian Academy of Sciences, 4 Kosygina Str.,
119991 Moscow, Russia
2 CNR, Istituto di Scienze e Tecnologie Chimiche “Giulio Natta”, Via De Marini 6, 16149 Genova, Italy;
giorgio.luciano@ge.ismac.cnr.it
3 Department of Materials, Federal University for Latin American Integration (UNILA), Silvio Américo Sasdelli
Avenue, 1842, Foz do Iguaçu-Paraná 85866-000, Brazil; hlornagj@ucs.br
4 Department of Physical Chemistry, Faculty of Chemical Technology, University of Pardubice, Studentská 95,
CZ-53210 Pardubice, Czech Republic; roman.svoboda@upce.cz
5 Department of Chemistry, University of Alabama at Birmingham, 901 S. 14th Street,
Birmingham, AL 35294, USA
* Correspondence: n.v.muravyev@ya.ru (N.V.M.); vyazovkin@uab.edu (S.V.)
Abstract: Artificial neural networks (ANNs) are a method of machine learning (ML) that is now
widely used in physics, chemistry, and material science. ANN can learn from data to identify
nonlinear trends and give accurate predictions. ML methods, and ANNs in particular, have already
demonstrated their worth in solving various chemical engineering problems, but applications in
pyrolysis, thermal analysis, and, especially, thermokinetic studies are still in an initiatory stage. The
present article gives a critical overview and summary of the available literature on applying ANNs
in the field of pyrolysis, thermal analysis, and thermokinetic studies. More than 100 papers from
these research areas are surveyed. Some approaches from the broad field of chemical engineering are
discussed as the venues for possible transfer to the field of pyrolysis and thermal analysis studies
in general. It is stressed that the current thermokinetic applications of ANNs are yet to evolve
significantly to reach the capabilities of the existing isoconversional and model-fitting methods.
Keywords: artificial neural networks; conversion degree; kinetics; machine learning; pyrolysis;
thermal analysis
1. Introduction
Machine learning has found a number of applications in chemistry [1–3] and material
science [4–6]. Among the supervised machine learning methods, the artificial neural
networks (ANN) are most popular. Artificial neural networks are the models inspired
by biological neural systems such as the human brain [7]. ANN represents the oriented
graphs with nodes called by the same analogy with brain as “neurons”. The key features of
the neural network are its topology (number of layers and neurons in it) and the strength
of connections between the neurons (defined by mathematical weights). The benefits of
ANN include the capability to approximate any continuous non-linear function, and high
tolerance to noisy or missing data [8,9]. Thanks to these properties, the neural networks
have found their use in vastly different fields, e.g., image classification [10], predicting
combustion instability [11], or impact sensitivity of energetic materials [12]. Although
ANNs have been extensively applied in chemical engineering since the 1990s [13–15], their
usage in thermal analysis has commenced later [9,16], and most of the studies emerged
only recently [17–21]. The present review aims to summarize such studies and to assess
some future prospects.
https://www.mdpi.com/journal/molecules
Molecules 2021, 26, x FOR PEER REVIEW 2 of 26

Molecules 2021, 26, 3727 2 of 25

Thermal analysis (TA) is defined as “the study of the relationship between a sample
property and its temperature as the sample is heated or cooled in a controlled manner”
Thermal analysis (TA) is defined as “the study of the relationship between a sample
[22]. In this review, we focus mainly on applications of neural networks to thermal anal-
property and its temperature as the sample is heated or cooled in a controlled manner” [22].
ysis
In and
this thermokinetic
review, we focusstudies.mainly Additionally,
on applications as ofone can see
neural from the
networks to literature survey,
thermal analysis
typical
and applicationsstudies.
thermokinetic involveAdditionally,
pyrolysis as aasthermal one canprocesssee from considered
the literature mostsurvey,
commonly typicalby
researchers. To keep the scope of review concise,
applications involve pyrolysis as a thermal process considered most commonly by re-the works on prediction of yield of py-
rolysis product
searchers. To keep (regarded
the scopeasofareview sort ofconcise,
TA experiment)
the worksare on included,
predictionwhile reports
of yield on the
of pyrolysis
selectivity and kinetics of pyrolysis gas products quantified
product (regarded as a sort of TA experiment) are included, while reports on the selectivity with ANNs [23–25] are not
covered
and kineticsin this review. gas products quantified with ANNs [23–25] are not covered in
of pyrolysis
The
this review. most common methods of thermal analysis are differential scanning calorimetry
(DSC) and thermogravimetric
The most common methods analysis
of thermal (TGA) normally
analysis applied under
are differential linearcalorimetry
scanning rising tem-
perature or constant temperature conditions. DSC produces
(DSC) and thermogravimetric analysis (TGA) normally applied under linear rising temper- a differential signal, i.e., the
heat flow data, that is generally assumed to be proportional
ature or constant temperature conditions. DSC produces a differential signal, i.e., the to the conversion rate ⁄𝑑𝑡.
𝑑𝛼heat
This data,
flow assumption ignores the
that is generally thermalto
assumed inertia term in the to
be proportional heat
theflow that is a rate
conversion potential
dα/dt.source
This
of a systematic
assumption errorthe
ignores in kinetic
thermal evaluations
inertia term [26]. However,
in the heat flowthisthat
error is tends to be negligible
a potential source of
awhen using smaller
systematic error inmasses
kineticand slower heating
evaluations rates [27].this
[26]. However, In turn,
errorTGA tends data,
to berepresenting
negligible
the mass
when using change
smaller of the
massessample,
and is an integral
slower heating signal, and its
rates [27]. In normalized
turn, TGA data, change yields the
representing
the mass change of the sample, is an integral signal, and its normalized change yields and
conversion degree α. The resulting conversion degree plotted against temperature the
time is also degree
conversion known α. as the
Thetransformation–time–temperature
resulting conversion degree plotted (TTT) dataset.
against With this data
temperature and
array,
time isthe
alsothermokinetic
known as theanalysis is then performed. The procedure
transformation–time–temperature relies onWith
(TTT) dataset. the following
this data
basic equation
array, [28]:
the thermokinetic analysis is then performed. The procedure relies on the following
basic equation [28]: 𝑑𝛼
dα𝑑𝑡 = 𝑘(𝑇)𝑓(𝛼), (1)
= k ( T ) f ( α ), (1)
The multiplicand in Equation (1),dti.e., the rate constant, usually assumes an Arrhenius
formThe 𝑘(𝑇) multiplicand
= 𝐴𝑒𝑥𝑝(−𝐸in Equation (1), i.e., the rate constant, usually assumes an Arrhenius
𝑎 ⁄𝑅𝑇 ) with Ea being the activation energy and A the preexponen-
form k T = Aexp E /RT
tial factor. The multiplier in)Equation
( ) (− a with Ea being (1) isthetheactivation and A the
energy Various
reaction model. preexponential
approaches have
factor. The multiplier in Equation (1) is the reaction
been proposed for thermokinetic analysis, the most effective being the model-fittingmodel. Various approaches have been
proposed
[29,30] and forthe
thermokinetic
isoconversional analysis,
[31,32] thetechniques.
most effective Thebeing the model-fitting
obtained kinetic parameters[29,30] gen-
and
the
erallyisoconversional
serve for obtaining [31,32]mechanistic
techniques. insights The obtained into the kinetic parameters
processes [33,34], generally serve for
or for predicting
obtaining mechanistic
their thermal behavior [35]. insights into the processes [33,34], or for predicting their thermal
behavior [35].
From the conceptual point of view, the traditional thermokinetic analysis, i.e., Equa-
tion From
(1) with thethe
conceptual point of(A,
kinetic triplet view,
Ea, the traditional
𝑓(𝛼)), thermokinetic
is a mechanistic, analysis, i.e., Equation
phenomenological, or para-(1)
with the kinetic triplet (A, E a , f ( α ) ), is a mechanistic,
metric modeling. The models of this type are known as white-box models, in contrast to phenomenological, or parametric
modeling. The models of this type are known as white-box models, in contrast to the
the data-driven or nonparametric black-box modelling (Figure 1). The papers discussed
data-driven or nonparametric black-box modelling (Figure 1). The papers discussed in
in the present review are primarily concerned with the latter approach. To give a simple
the present review are primarily concerned with the latter approach. To give a simple
example of a black box model, let us imagine a neural network that is trained to output
example of a black box model, let us imagine a neural network that is trained to output
the conversion rate data once fed with temperature and time values. This ANN is readily
the conversion rate data once fed with temperature and time values. This ANN is readily
applicable and does not require knowledge of the process kinetics. If carefully designed
applicable and does not require knowledge of the process kinetics. If carefully designed
and properly trained, the network offers good descriptive accuracy within the domain of
and properly trained, the network offers good descriptive accuracy within the domain
the supplied process data. Main limitations of the black-box models are the uncertain in-
of the supplied process data. Main limitations of the black-box models are the uncertain
terpretation [36] and poor generalization ability [37]. To combine the benefits of white-
interpretation [36] and poor generalization ability [37]. To combine the benefits of white-
and black-box approaches, hybrid semi-parametric modeling is proposed [38] (Figure 1).
and black-box approaches, hybrid semi-parametric modeling is proposed [38] (Figure 1).
The application of hybrid models will be addressed in detail in Section 4 of this review. It
The application of hybrid models will be addressed in detail in Section 4 of this review.
is isworth
It worthmentioning
mentioningthat thatsome
someof ofthethe further-discussed
further-discussed ANNs ANNs are trained on
are trained on synthetic
synthetic
data generated with Equation (1) or give as an output
data generated with Equation (1) or give as an output the kinetic parameters of Equation the kinetic parameters of Equation
(1). Such models cannot be regarded as authentic
(1). Such models cannot be regarded as authentic black-box models because they eitherblack-box models because they either
implicitly or explicitly utilize a phenomenological
implicitly or explicitly utilize a phenomenological model. model.

Figure 1.
Figure 1. Three
Three conceptual
conceptual approaches
approaches to
to chemical
chemical engineering
engineeringmodeling.
modeling.
Molecules 2021, 26, x FOR PEER REVIEW 3 of 26
Molecules 2021, 26, 3727 3 of 25

The present review briefly surveys the conceptual details of ANNs, and summarizes
theirThe present review
applications briefly thermal
in pyrolysis, surveys analysis,
the conceptual details
and kinetic of ANNs,
studies. and summarizes
A strong emphasis is
their applications in pyrolysis, thermal analysis, and kinetic studies. A strong emphasis
put on the kinetic applications as this is a quickly growing field that needs special is
atten-
put on the kinetic applications as this is a quickly growing field that needs special attention
tion for proper development. Figure 2 schematically shows the blocks that represent major
for proper development. Figure 2 schematically shows the blocks that represent major
topics identified via analysis of the literature. The article is roughly structured according
topics identified via analysis of the literature. The article is roughly structured according
to these topics that are presented as the article sections. Note that the divisions between
to these topics that are presented as the article sections. Note that the divisions between
the sections are rather conditional and that the topics covered oftentimes overlap with
the sections are rather conditional and that the topics covered oftentimes overlap with
other chemical engineering or chemometric problems. The review is intended to reflect
other chemical engineering or chemometric problems. The review is intended to reflect
the existing state of affairs in the area of the aforementioned applications and to provide
the existing state of affairs in the area of the aforementioned applications and to provide
certain critique, and to highlight some desired directions for future research.
certain critique, and to highlight some desired directions for future research.

Figure2.2. Logical
Figure Logical scheme
scheme of
of the
the present
present review.
review. Several
Several aspects
aspects of
of using
using artificial
artificial neural
neural networks
networks (ANN)
(ANN) for
forthermal
thermal
analysis (TA) studies are indicated to be discussed in main text.
analysis (TA) studies are indicated to be discussed in main text.

2. Technical Details behind ANNs

2. Technical Details behind
Before discussing ANNs
general aspects of the ANNs implementation, we should recognize
that Before
there isdiscussing
no single general aspectstool
mathematical of the ANNs
that implementation,
resolves all TA-specific weproblems.
should recognize
ANN is
that there is no single mathematical tool that resolves all TA-specific
only one of several machine learning methods and its accuracy for specific applications problems. ANN is onlyis
one of several
hardly assessed machine
a priori learning methodsisand
(this situation knownits accuracy for specific
in mathematics as aapplications is hardly
“no free lunch” the-
assessed
orem [39]).a priori (this situation
However, for some is known in mathematics
applications the ANNasmodel a “no has
free been
lunch”compared
theorem [39]).
with
However,
other machinefor some applications
learning the ANN
approaches, model has
e.g., random been[40],
forest compared
gradient with other machine
boosting machine
learning
[40], andapproaches,
it has been founde.g., random
that ANN forest [40], gradient
performs better. boosting machine [40], and it has
been Several
found that ANN
types performs
of ANNs are better.
used in chemical engineering studies (e.g., see review by
Several types of ANNs
Ali et al. [41]). Nonetheless, almost are used allinpapers
chemical engineering
considered studies
in the present (e.g., see review
review by
utilize the
Ali et al. [41]). Nonetheless, almost all papers considered
same type of ANN known as feed-forward neural network, which is briefly described in the present review utilize the
same
below.type
More of practical
ANN known details asare
feed-forward neuralfrom
readily available network, which issources
monographic briefly [42,43].
described
below. ANNMoreispractical
a paralleldetails are readily
processing available
system from monographic
of interconnected sources [42,43].
computational elements,
ANN is a parallel processing system of interconnected
called neurons (Figure 3). The neurons are arranged in several layers—input, hidden,computational elements, called
and
neurons (Figure 3). The neurons are arranged in several layers—input,
outer layers. The topology of ANN is encoded numerically as follows, e.g., 25-15-10 hidden, and outer
de-
layers.
notes theTheneural
topology of ANN
network is encoded
with 25 neurons numerically
in the input as follows,
layer, 15e.g., 25-15-10
neurons indenotes
the hidden the
neural network
layer, and with 25
10 neurons inneurons
the output in the input
layer. Note layer,
that 15mostneurons
studiesinreviewed
the hidden layer,
adopt theand
ar-
10
chitecture with a single hidden layer, rarely two hidden layers are used [20,44], andwith
neurons in the output layer. Note that most studies reviewed adopt the architecture only
afor
single hidden layer,
exceptionally rarelythermal
complex two hidden layers
profiles theare used [20,44],
authors has had andtoonly for exceptionally
introduce even more
complex thermal profiles the authors has had to introduce even more hidden layers [45].
hidden layers [45]. Connections between the neurons are displayed as lines in Figure 2,
Connections between the neurons are displayed as lines in Figure 2, each connection
each connection has its own strength value called weight. Output of the neuron is deter-
has its own strength value called weight. Output of the neuron is determined by each
mined by each layer’s transfer functions (activation function). While in most studies the
layer’s transfer functions (activation function). While in most studies the hyperparameters
hyperparameters (parameters used to control the learning process) and parameters of
(parameters used to control the learning process) and parameters of ANN are tuned
ANN are tuned manually, there are some techniques to make the procedure more rigor-
manually, there are some techniques to make the procedure more rigorous and human-
ous and human-independent. In particular, the differential evolution algorithm can be
independent. In particular, the differential evolution algorithm can be utilized to optimize
utilized to optimize the number of hidden layers, number of neurons in it, weights, biases,
the number of hidden layers, number of neurons in it, weights, biases, and activation
and activation functions [46].
functions [46].
Let us consider ANN in more detail. The output of i-th neuron in (L+1)-th layer is
Let us consider ANN in more detail. The output of i-th neuron in (L + 1)-th layer is
denoted as 𝑦L𝑖𝐿+1 . This value is comprised of a specific constant value called bias 𝑏 𝐿+1 , and
denoted as yi +1 . This value is comprised of a specific constant value called bias biL𝑖+1 , and
Molecules 2021, 26,
Molecules 2021, 26, 3727
x FOR PEER REVIEW 44 of
of 25
26

NLL signals
signals from
from previous
previous layers,
layers, amplified
amplified or
or weakened
weakened by
by the
the corresponding
corresponding coefficients,
coefficients,
called weights w 𝑤ijL𝑖𝑗𝐿+1
+1 [19]:
[19]:
𝑁
𝑦𝑖𝐿+1 =𝑓(∑ 𝐿 𝐿+1 𝐿
N𝑗=1 𝑤𝑖𝑗 𝑦𝑗 + 𝑏𝑖
𝐿+1
). (2)
yiL+1 = f ∑ j=L1 wijL+1 y Lj + biL+1 . (2)
Among a broad variety of the activation functions [42,43], those of sigmoidal type
𝑓(𝑢) Among
= 𝐿𝑜𝑔𝑠𝑖𝑔(𝑢) a broad = 1/(1 + 𝑒 −𝑢
variety of), thesymmetric
activationsigmoidal form 𝑓(𝑢)
functions [42,43], those= 𝑇𝑎𝑛𝑠𝑖𝑔(𝑢)
of sigmoidal = −1type +
2/(1 + 𝑒 −2𝑢
) − u
f (u) = Logsig
(u) = 1/(1 + e ), symmetric sigmoidal form f (u) = Tansig(u) = −1 +
have been used most frequently. Thus, the ANN model transforms the input
vector
2/ 1 +(xe− 2u have
1, x2, …, xi)been
to the output
used mostvector (y1, y2,Thus,
frequently. … yj).theBefore
ANNusing,model ANN is trained
transforms theon the
input
[(x1, y1),(x…,
vector 1 , x(x
2 , n., .y. n,)]xdata,
i ) to theby output
modifying vectorthe(yweights
1 , y2 , . .of
. yconnections between
j ). Before using, ANN theisneurons
trained
on the [(x1 , ythe
to minimize 1 ), .prediction
. . , (xn , yn )] data,
error by modifying
(supervised the weights
learning). Although of connections
more efficient between
and mod- the
neurons to minimize
ern algorithms (e.g., the prediction
Stochastic error (supervised
Gradient learning). Moment
Descent or Adaptive AlthoughEstimation)
more efficient are
and modern algorithms (e.g., Stochastic Gradient Descent
used broadly in ANN training, the application considered in the present review appear to or Adaptive Moment Estimation)
are used broadly
be limited primarily in ANN to the training, the application considered
Levenberg–Marquardt in the present review
or error backpropagation appear
algorithms.
to
Thebekeylimited
problem primarily with the to the Levenberg–Marquardt
optimization or error backpropagation
is that it can converge to local minima. Some algorithms.
mod-
The key problem
ifications for the common with theprocedures
optimization areisconstantly
that it canproposed
convergetotoimprove
local minima. Some
the computa-
modifications
tional time andfor the common
robustness. procedures
For example, antare constantly
colony proposed
optimization (ACO)to improve
algorithm theiscom-
sug-
putational time and robustness. For example, ant colony
gested [47] to improve the convergence of backpropagation technique, and it is shown optimization (ACO) algorithm
is suggested
that it slightly[47] improvesto improve the convergence
the performance of ANN of backpropagation
[48]. To evaluate how technique,
well ANN and pre-
it is
shown that it slightly improves the performance of ANN
dicts in the domains not used for training, the data are usually split into the training and [48]. To evaluate how well ANN
predicts
validation insets.
the domains
The split not ratioused
shouldfor generally
training, the data on
depend aretheusually
qualitysplit
andinto the training
structure of the
and validation sets. The split ratio should generally
data. However, our analysis of the literature concerned with the present review topic depend on the quality and structure
sug-
of the data. However, our analysis of the literature concerned
gests that the typical ratio for sizes of the training and validation sets is ~70% and ~30%, with the present review
topic suggestsThe
respectively. thatrealistic
the typical ratio forlimits
confidence sizes offortheANNtraining and validation
predictions sets is ~70%
can be generated asand
de-
~30%, respectively. The realistic confidence limits for
scribed elsewhere [49]. Ensemble learning techniques, like stacking of several ANNs, ANN predictions can be generated as
described elsewhere [49]. Ensemble learning techniques,
boosting or bagging, have been suggested as a viable methods to further improve the gen- like stacking of several ANNs,
boosting
eralizationorability
bagging, [50].have been suggested as a viable methods to further improve the
generalization ability [50].

Figure Schemeofofthe
3. Scheme
Figure 3. theartificial
artificial neural
neural network
network with
with indicated
indicated layers,
layers, weights,
weights, functions
functions [51]. [51].
Reprinted with permission from Taylor & Francis.
Reprinted with permission from Taylor & Francis.

Optimal ANN architecture is defined by learning quality and generalization ability.

Optimal ANN architecture is defined by learning quality and generalization ability.
Both parameters rely on the data, which should be representative and big enough in
Both parameters rely on the data, which should be representative and big enough in size.
size. However, as the studies considered below mostly involve the time-consuming and
However, as the studies considered below mostly involve the time-consuming and labor-
labor-intensive measurements, the size of available data is usually small, especially when
intensive measurements, the size of available data is usually small, especially when com-
compared with typical applications of machine learning. Just to give an example of typical
pared with typical applications of machine learning. Just to give an example of typical
number of learning data, 50,000 images are used to train ANN to recognize the handwritten
Molecules 2021, 26, 3727 5 of 25

numbers from 1 to 10 [52]. For the thermal behavior data, the amount of input data is
defined by a particular experimental setup and sampling frequency, but usually is far less
than the number above. To give an estimate of how much data is needed, we notice that
the theoretical estimates suggest that the number of learning patterns necessary to attain
the sufficient generalization ability should be ten times larger than the total number of
weights in ANN [53,54]. The problem of sparse data can be alleviated by the generation of
the artificial raw thermal analysis data [21], as discussed in the following section.
An important aspect of the application of ANNs to the TA-generated data is the
representation of the input data. TA data used for thermokinetic analysis are usually in
the form of several sets of three-dimensional data (conversion, temperature, time). These
data should be transformed to decrease the dimensionality of the information being fed
to the neurons of the input layer. The simplest way of doing this is to limit input to a
single experiment and input one of the dependent parameters, e.g., time for isothermal
runs [17], and train ANN to output the conversion data (note, that utilization of data
corresponding only to a single run is not recommended from the kinetic point of view [28],
although in some specific cases the approach could work). Raw thermal analysis signal can
be considered as the output, while the temperature is the independent variable and each
heating rate is considered separately [20,55]. A set of temperature values corresponding
to the specific conversion degrees is used as inputs, and the single row of input vector is
formed from several experiments performed at different heating rates [21,56]. This thin-out
procedure should retain the kinetic information, i.e., have considerable number of points,
and contain the regions with critical information [9,16] (peak, inflexion, etc.).
Some more advanced methods for mapping the multi-dimensional thermal analysis
data for the input layer of neural network still need to be developed. Muravyev and
Pivkina [21] propose using principal component analysis [57–59] or introducing additional
hidden layers for this purpose. To the best of authors’ knowledge, no more advanced
methods of introduction of TA data to ANN have been proposed so far. Possibly, some
methods from other fields can be borrowed, e.g., that employed for spectral dimensionality
reduction in spectroscopy and signal analysis [60].
Finally, reporting of the optimized architecture of the ANN model is strongly recom-
mended. Only a few examples [19,61,62] are found in the reviewed literature that provide
the network biases and weights. As a guiding principle, one must provide enough detail
on the ANN architecture, so that others can reproduce the computations. This principle is
commonly accepted in reporting experimental studies. It is only prudent to follow it in
publishing ANN work.

3. Prediction of Conversion Data (Single Value, Whole Curve)

There is a number of papers [63–66] on the application of ANNs for obtaining the
total conversion (e.g., mass loss by TGA) using experimental conditions as an input. One
of the first papers on the topic has been by Carsky and Kuwornoo [67] who used ANN
to predict the main quantities for coal pyrolysis: tar, volatiles, and char yields. Table 1
shows the input parameters, which those authors have compiled from the literature on
pyrolysis of several varieties of coals in fluidized bed reactor, spouted bed reactor, hot-rod,
and wire-mesh reactors (since the experiments comprise the heating and isothermal hold
steps, we consider them as TA data). An important aspect of the study is that the input
parameters are in most cases not distributed uniformly or continuously. Thus, the authors
have used, additionally, the fuzzy modeling to transform the particle holding time. For
other variables, e.g., particle size, a qualitative classification (fine, coarse) was used instead
of unevenly distributed numerical values. The optimized neural network (47-16-3) has
described well the experimental tar, volatiles, and char yields, as is shown by root mean
square errors of 4.2, 4.9, and 6.4%, respectively.
Molecules 2021, 26, 3727 6 of 25

Table 1. The list of the parameters used to model the coal pyrolysis, and some details on their incorporation to ANN [67].
Reprinted with permission from Elsevier.

Parameter Format Transformation Input/Output

Particle holding time Fuzzy class membership (s) Logical (0/1) Input
Particle holding time Fuzzy class membership (min) Logical (0/1) Input
Linear, Inverse function, Inverse
Freeboard residence time
Numerical (s) fourth root function, Input
of volatiles
Fuzzy center (0) function
Temperature Numerical (◦ C) Linear, Tanh, Square root function Input
Linear, Inverse fourth root function,
Time of preheating of particles Numerical (s) Fuzzy center (0) function, Input
Inverse function
Linear, Inverse fourth root function,
Rate of heating Numerical (K/s) Input
Inverse function
Secondary reaction Qualitative (none, small, Enumerated strings:
Input
in freeboard medium, great) small, medium, great
Qualitative (small,
Secondary reaction in bed Enumerated strings: small, great Input
medium, great)
Reactor type String (flb, spb, hr, wm) 1 Enumerated strings: flb, spb, hr, wm Input
Particle size Qualitative (fine, coarse) Enumerated strings: fine, coarse Input
Enumerated strings:
Concentration of coal particles Qualitative (low, medium, high) Input
low, medium, high
Qualitative (shallow, deep, Enumerated strings: shallow, deep,
Bed depth medium, mono-layer, medium, mono-layer, Input
mono-layer-shallow) mono-layer-shallow
Linear, Inverse fourth root function,
Sample size Numerical (g of coal) Input
Inverse function

2 Linear, Inverse fourth root function,

Coal: wt% ash (d.b.) Numerical (% db) Input
Inverse function

3 Linear, Fourth power function,

Coal: wt% volatiles (d.a.f.) Numerical (% daf) Input
Exponential
Linear, Inverse fourth root function,
Carrier gas velocity Numerical (m/s) Input
Inverse function
Yield of tar Numerical (wt%) Linear Output
Yield of volatiles Numerical (wt%) Inverse function Output
Yield of char Numerical (wt%) Inverse function Output
1 Designation of reactor types: flb = fluidized bed reactor, spb = spouted bed reactor, hr = hot-rod fixed bed reactor, wm = wire-mesh
reactor. 2 d.b.—dry basis value (except all moisture). 3 d.a.f.—dry ash free value (except all moisture and ash).

Thus, originally, this approach has been proposed to derive a single final value of
conversion. In more recent studies, the idea has been extended to employ ANNs for pre-
dicting the conversion values at various moments, i.e., a TA signal itself. Such studies differ
drastically in experimental systems (Tithonia diversifolia weed biomass [63], low-density
polyethylene [64], high-density polyethylene [68], safflower seed press cake [65], durian
rinds [69], rape straw [70], coal gangue and peanut shell [71], pet coke [72], sewage sludge
and peanut shell [73], sewage sludge and coffee grounds [74], vegetable fibers [75], rice
husk and sewage sludge [45], sludge, watermelon rind, corncob, and eucalyptus [76],
Sargassum sp. seaweed [77], cotton cocoon shell, tea waste, and olive husk [66], mechanoac-
tivated coals [78], cattle manure [79], lignocellulosic forest residue and olive oil residue [80],
cotton cocoon shell, fabricated tea waste, olive husk, and hazelnut shell [81]) and in some
minor details, but the general concept remains the same. Thus, we illustrate it with the
Molecules 2021, 26, 3727 7 of 25

Molecules 2021, 26, x FOR PEER REVIEW 7 of 26

study by Kataki et al. [63], who used a neural network model (4-14-1) to predict the product
yield for the pyrolysis of dried weed biomass. The four input parameters have been the
been the pyrolysis
pyrolysis temperature
temperature (during the (during
30 min theisothermal
30 min isothermal
segment), segment), heating
heating rate usedrate used
to attain
to attain this temperature, nitrogen flow rate, and particle size of
this temperature, nitrogen flow rate, and particle size of the specimen. Thirty pyrolysis the specimen. Thirty
pyrolysis
experimentsexperiments with varying
with varying experimentalexperimental
conditions conditions
have been have been designed
designed and per-
and performed,
formed,
and the and
target the target(yield
output output of (yield of thehas
the bio-oil) bio-oil)
been has been determined.
determined. The optimized
The optimized ANN has
ANN
demonstrated a better accuracy than the response surface method used as an used
has demonstrated a better accuracy than the response surface method as an
alternative.
alternative. Finally, ANN has been used to design the experimental
Finally, ANN has been used to design the experimental conditions for maximizing conditions for max-
the
imizing the output quantity, and the result has agreed closely with
output quantity, and the result has agreed closely with the neural network prediction. the neural network
prediction.
Although the Although the inner relationships
inner relationships of ANN are of ANN are not some
not known, known, some insights
insights can be
can be gained
gained by looking
by looking at the at the relative
relative importance
importance of theofinput
the input descriptors.
descriptors. FigureFigure 4 shows
4 shows thatthat
the
the pyrolysis
pyrolysis temperature
temperature is main
is the the mainfactorfactor
(more(more
thanthan
30% 30% of relative
of relative importance)
importance) fol-
followed
lowed by heating
by heating rate flow
rate (25%), (25%), flow
rate rateand
(22%), (22%), and size
particle particle size
(18%). (18%).
The The discussed
discussed approach ap-can
proach can be considered as a sort of experimental design and can
be considered as a sort of experimental design and can be definitely used for industrial be definitely used for
industrial thermal processes
thermal processes to optimize to the
optimize
singlethe single
target target (e.g.,
property property
yield(e.g., yield of product).
of product).

Theexample
Figure4.4.The
Figure example of
of relative
relative importance analysis
analysis for
forinput
inputdescriptors
descriptorsofoftrained
trainedneural
neuralnetwork
net-
model [63]. Reprinted by permission from Springer Nature.
work model [63]. Reprinted by permission from Springer Nature.

Theidea
The ideafrom
from the
the above-mentioned
above-mentionedstudiesstudiescancan be
be expanded
expanded further.
further. ANNANN can can be
be
trained to predict the conversion degree (or the primary TA signal)
trained to predict the conversion degree (or the primary TA signal) based on temperaturebased on temperature
and time.
and time.This Thisapproach
approach cancan
bebe classified
classified as as a nonlinear
a nonlinear fit the
fit of of the
TA TA data,
data, because
because no
extrapolation is performed, and all data are within the experimental domain space.space.
no extrapolation is performed, and all data are within the experimental domain The
The graphical
graphical representation
representation of the workflow
of the workflow from the from thebystudy
study Bezerra by et
Bezerra
al. [20] et al. [20]in
is shown is
shown in Figure 5a. It is rather a general approach that predicts the thermal
Figure 5a. It is rather a general approach that predicts the thermal signal from a principal signal from
a principal(temperature)
descriptor descriptor (temperature)
and another and anotherwhich
parameter, parameter,
defines which defines nonisother-
the specific the specific
nonisothermal measurement, the heating rate. Figure 5b,c show
mal measurement, the heating rate. Figure 5b,c show the good quality fit via neural the good quality fit
net-
via neural network models applied to complex thermally-induced
work models applied to complex thermally-induced processes. Figure 5b illustrates the processes. Figure 5b
illustrates theTGA
experimental experimental
signal forTGA signal for nonisothermal
nonisothermal pyrolysis
pyrolysis of a carbon of a carbon
reinforced reinforced
fiber compo-
fiber composite along with its fit via the ANN model (2-21-21-1). Figure 5c represents the
site along with its fit via the ANN model (2-21-21-1). Figure 5c represents the isothermal
isothermal data for iron oxide reduction process [19] fitted with a neural network (3-20-1).
data for iron oxide reduction process [19] fitted with a neural network (3-20-1). In both
In both cases, ANN offers an excellent fit of experimental data.
cases, ANN offers an excellent fit of experimental data.
As one may notice, the ANN used to obtain the data shown in Figure 5b contains two
As one may notice, the ANN used to obtain the data shown in Figure 5b contains two
hidden layers, whereas the one used to fit the data from Figure 5c contains only a single
hidden layers, whereas the one used to fit the data from Figure 5c contains only a single
hidden layer of neurons. The selection of complexity of the ANN topology is, thus, forced
hidden layer of neurons. The selection of complexity of the ANN topology is, thus, forced
by the data. This observation is in line with the findings by Cepeliogullar et al. [44]. They
by the data. This observation is in line with the findings by Cepeliogullar et al. [44]. They
have studied the pyrolysis of refuse-derived fuel and analyzed in detail the impact of the
have studied the pyrolysis of refuse-derived fuel and analyzed in detail the impact of the
ANN architecture parameters. Due to the complex thermal behavior of the investigated
ANN architecture parameters. Due to the complex thermal behavior of the investigated
highly heterogeneous samples, only after introduction of the second hidden layer of
highly heterogeneous samples, only after introduction of the second hidden layer of neu-
neurons an adequate fitting quality is achieved. In addition to the number of neurons in
rons
bothan adequate
hidden fitting
layers, the quality
authorsisconsidered
achieved. In addition
various to the number
activation functions. of neurons in both
For the optimal
hidden layers, the authors considered various activation functions.
neural network, the prediction ability was tested by comparison with direct experimentFor the optimal neural
network, the prediction ability was tested by comparison with direct
not used in the model development. The predicted mass loss curve for heating rate experiment not usedof
in the model
− 1 development. The predicted mass loss curve for heating
25 K min closely agrees with the experiment. Note, however, that ANN has been trained rate of 25 K min −1

closely agrees with the experiment. Note, however, that ANN has been trained with data
Molecules 2021, 26, x3727
FOR PEER REVIEW 8 of8 26
of 25

acquired
with dataatacquired
5, 10, 15,
at 20, 30,15,
5, 10, 40,20,
and
30, 50
40,Kand
min
50 ,Kthus,
−1
min−this prediction
1 , thus, is still within
this prediction that
is still within
domain space and it is, in fact, a sort of interpolation.
that domain space and it is, in fact, a sort of interpolation.

Figure Typicalapproach
5. Typical
Figure 5. approach to to thermal
thermal signal
signal fit with
fit with ANN ANN model:
model: (a) the(a) the workflow
workflow [20],
[20], and and examples
examples of use
of use (b,c). Noni- (b,c).
Nonisothermal − 1
sothermal data data (b) corresponds
(b) corresponds to pyrolysis
to the the pyrolysis of carbon
of carbon reinforced
reinforced carbon
carbon composite
composite (heating
(heating rate
rate of of
1010KK min) [20].
min −1 ) [20].
Isothermal data (c) represents
represents the
the iron
iron oxide
oxide reduction
reduction(T (T==880 ◦ C,environmental
880°C, environmentalgasgas90%
90%NN 2 + +
2 10%
10% CO)
CO)[19]. Reprinted
[19]. Reprinted
with permission
with permission from
from Elsevier.
Elsevier.

Burgazetetal.al.[82]
Burgaz [82] have
have applied
applied thethe neural
neural network
network models
models to directly
to directly predictpredict
several sev-
thermal analysis signals (DSC, TGA, and dynamic mechanical analysis, DMA) forfor
eral thermal analysis signals (DSC, TGA, and dynamic mechanical analysis, DMA)
poly(ethylene oxide)/clay
poly(ethylene oxide)/claynanocomposites.
nanocomposites.The Theinput
inputparameters
parameters are are the
the amount
amount of of clay
clay in
in the sample and temperature, other details are summarized in Table 2. Authors note thatthe
the sample and temperature, other details are summarized in Table 2. Authors note that
amount
the amountof data necessary
of data to attain
necessary goodgood
to attain performance
performanceis signal-dependent,
is signal-dependent,i.e., iti.e.,
is smaller
it is
for DMA
smaller forthan
DMA forthan
the for
TGA theand
TGA DSCandsignals. The higher
DSC signals. complexity
The higher of the mass
complexity of theloss
mass and
heatand
loss flow signals
heat flow also forces
signals alsothe use the
forces of two
usehidden layers instead
of two hidden of only of
layers instead one forone
only DMA. for
DMA.
Table 2. Details of the neural network
From the models for prediction
research discussed of in various thermal
this section, we signals for poly(ethylene
can conclude that ANNsoxide)/clay
can success-
nanocomposites [82]. Reprinted with permission from Elsevier.
fully fit various thermal signals. The topology of the ANN model and other parameters
ANN and Training
depend on the amount of learning data, the selected input parameters, but most pro-
DMAon the inherent complexity
foundly DMA of data. The examplesTGA available so far do DSC not include
Parameters
Input data some testing on
Temperature and clay wt%distant extrapolation accuracy
Temperature and clay wt% of the neural networks
Temperature and clay wt% output. Research
Temperature and clay wt% in
Output data
this direction
E’
would certainly be of
tanδ (E”/E’)
value. In all, due to its excellent
Weight loss
interpolation accuracy,
Heat flow
this particular approach (conversion prediction from the experimental conditions) already
Training data 193 193 406 565
has a plenty of industrial applications.
Testing data 47 47 45 141
Number of hidden layers 1 1 2 2
Activation functions Tansig-Purelin Tansig-Purelin Tansig-Logsig-Purelin Tansig-Logsig-Purelin
Number of neurons in
9 9 4–3 5–4
hidden layer
Number of iterations 288 152 48 65
Regression (R2) 0.9997 0.9982 0.99994 0.99985
Molecules 2021, 26, 3727 9 of 25

From the research discussed in this section, we can conclude that ANNs can success-
fully fit various thermal signals. The topology of the ANN model and other parameters
depend on the amount of learning data, the selected input parameters, but most profoundly
on the inherent complexity of data. The examples available so far do not include some
testing on distant extrapolation accuracy of the neural networks output. Research in this
direction would certainly be of value. In all, due to its excellent interpolation accuracy, this
particular approach (conversion prediction from the experimental conditions) already has
a plenty of industrial applications.

4. Advancements in Prediction of Conversion Data with ANNs (Moving Window,

Hybrid Models)
It appears that chemical engineering has several approaches to the ANN applications
that can be transferred to pyrolysis and thermal analysis problems. Note that publications
discussed in the previous section do not use the conversion data as input, but using
such input in ANNs with moving time window is highly beneficial. Figure 6 illustrates
schematically the concept [54]: the input data include the temperature (pressure) data and
the concentration values from previous time steps and the output data are the concentration
of the reaction products at this particular time period. The depth of the time shift and
the sampling time become important parameters of the model. The neural model with
moving window outperforms the traditional neural network as shown on the data for
hydrogenation of 2,4-dinitrotoluene in stirred tank reactor [54]. The main drawback of this
Molecules 2021, 26, x FOR PEER REVIEW 10 of 26
kind of ANNs is the increased number of parameters that necessitates large quantities of
input data and results in considerable computational times [83].

Modelingofofthe
Figure 6.6.Modeling
Figure thereactor
reactorbybyneural
neural networks
networks with
with moving
moving time
time window
window [54].
[54]. X isXthe
is the
vec-vector
tor of input data, Y is vector of output data. Reprinted with permission from Elsevier.
of input data, Y is vector of output data. Reprinted with permission from Elsevier.

The utilization
The utilization ofof neural
neuralnetworks
networksprimarily
primarilyfor forprediction
predictionof ofthe
theconversion
conversionprofiles
profiles can
be considered
can be consideredas aassort of black-box
a sort of black-boxapproach. It has
approach. no need
It has no needof knowledge
of knowledge of the
of process
the
fundamentals.
process However,
fundamentals. alternatively
However, to this to
alternatively fully
thisblack-box approach,
fully black-box the hybrid
approach, approach
the hybrid
is proposed
approach [38], being[38],
is proposed a combination of traditional
being a combination modeling with
of traditional description
modeling by ANN of the
with description
by ANN of the unknown part of the process (e.g., the reaction kinetics). It is
unknown part of the process (e.g., the reaction kinetics). It is deemed that the latter techniquedeemed that
the latter technique
is beneficial is beneficial
as it affords as it affords
some degree of thesome degree of the
generalization of generalization
the results [54]. of We
the re-are not
sults
aware[54]. We are
of using thisnot aware of
approach tousing this approach
the specific to the
applications specific
(i.e., applications
to pyrolysis, (i.e.,analysis,
thermal to
pyrolysis, thermal analysis,
and thermokinetic studies) and thermokinetic
considered studies) review,
in the present consideredthus,inwe
themake
present
an review,
excursus in
thus, we make
chemical an excursus
engineering in chemical
to introduce engineering
the basic idea of thetomethod.
introduce the basic idea of the
method.
To exemplify the hybrid approach, let us consider the study of heterogeneous oxida-
tion of 2-octanol with nitric acid by Molga et al. [84] The chemical reaction was performed
in the reaction calorimeter (RC1, Mettler-Toledo) and during the experiment, a small
amount of samples were extracted and analyzed chemically. The adopted reaction mech-
Molecules 2021, 26, 3727 10 of 25

To exemplify the hybrid approach, let us consider the study of heterogeneous oxidation
of 2-octanol with nitric acid by Molga et al. [84] The chemical reaction was performed in
the reaction calorimeter (RC1, Mettler-Toledo) and during the experiment, a small amount
of samples were extracted and analyzed chemically. The adopted reaction mechanism
comprises the oxidation of 2-octanol (A) to 2-octanone (P) by action of nitrosonium ion (B),
and the subsequent oxidation with formation of carboxylic acids (X):

A + B → P + 2B, P + B → X. (3)

The neural network works here as a kinetic model: it gets the temperature and
concentrations of A, P, X, B as inputs and outputs the reactions rates for two reactions,
R1 , R2 . ANN is then used in combination with traditional mass balance and heat balance
equations, e.g.:
dn P
= ( R1 − R2 )(n A,0 + n N,0 ), (4)
dt
Molecules 2021, 26, x FOR PEER REVIEW (ni is molar amount of product i in the system, subscript N stands for the nitric acid, a
11 of 26
source of nitrosonium ion). That is, the hybrid model is built, where the system of equations
is solved with the kinetic data coming from the trained ANN (Figure 7).

Figure 7. Schematic diagram of the hybrid model of chemical reactor [84]. Reprinted with permission
Figure 7. Schematic diagram of the hybrid model of chemical reactor [84]. Reprinted with permis-
fromfrom
sion Elsevier.
Elsevier.

This hybrid approach is well developed in chemical engineering [85], but it has not
This hybrid approach is well developed in chemical engineering [85], but it has not
been yet applied systematically to the problems considered in the present review. One
been yet applied systematically to the problems considered in the present review. One of
of the examples of its application is the study by Guo et al. [86], where the hybrid model
the examples of its application is the study by Guo et al. [86], where the hybrid model was
was built for gas production rates in the course of coal pyrolysis. More research with this
built for gas production rates in the course of coal pyrolysis. More research with this ap-
approach would be welcome. For example, it should be useful for purposes of thermal
proach would be welcome. For example, it should be useful for purposes of thermal be-
behavior prediction [35]. The conversion rate term will be supplied by ANN from input
havior prediction [35]. The conversion rate term will be supplied by ANN from input Ti,
Ti , αi-1 and then the output dα/dt|i value will enter the conventional phenomenological
α i-1 and then the output dα/dt|i value will enter the conventional phenomenological
model, thermal balance equations for storage condition [87]:
model, thermal balance equations for storage condition [87]:
𝑑𝑇𝑠 𝑑𝑇𝑐 𝑑𝛼
𝑚𝑠 𝑐𝑝,𝑠 dT +s 𝑚𝑠 𝑐𝑝,𝑐 dT =c 𝑈𝑆(𝑇𝑒𝑛𝑣 − 𝑇𝑐 ) + 𝑚𝑠 ∆𝐻𝑑 , dα
ms c p,s 𝑑𝑡 + m c
s p,c𝑑𝑡 = US ( Tenv − Tc ) + m ∆H
s 𝑑𝑡 d , (5)
dt dt dt
where ΔHd is the enthalpy of decomposition reaction, U–the heat transfer coefficient, S–
where ∆H is the enthalpy of decomposition reaction, U—the heat transfer coefficient,
the area ofd a contact surface between sample and container, cp, m, and T stand for the
S—the area of a contact surface between sample and container, c , m, and T stand for the
isobaric heat capacity, mass, and temperature, correspondingly. pThe subscripts s, c, env.
isobaric heat capacity, mass, and temperature, correspondingly. The subscripts s, c, env.
refer to the sample, container, and environment parameters, respectively.
refer to the sample, container, and environment parameters, respectively.
There are other ways of combining traditional modeling with that by neural network
There are other ways of combining traditional modeling with that by neural net-
[85,88]. An example is a study by Curteanu et al. [46] where the conversion and molar
work [85,88]. An example is a study by Curteanu et al. [46] where the conversion and
weights (α, MW) for the free radical polymerization of styrene are obtained by various
molar weights (α, MW) for the free radical polymerization of styrene are obtained by
models as the functions of time, temperature, and concentration of the initiator. Specifi-
various models as the functions of time, temperature, and concentration of the initiator.
cally, the authors consider four situations: (1) serial, the process parameters are an input
Specifically, the authors consider four situations: (1) serial, the process parameters are
to the traditional model and its output (α, MW) is sent to the ANN model that gives the
an input to the traditional model and its output (α, MW) is sent to the ANN model that
final
givesvalues of (α,
the final MW);
values of(2)(α,parallel,
MW); (2) both, the traditional
parallel, both, themodel and ANN
traditional modelare fedANN
and with the
are
process parameters and give the (α, MW) values, these values
fed with the process parameters and give the (α, MW) values, these values are averagedare averaged to give theto
final result;
give the final(3)result;
serial-parallel, the ANNthe
(3) serial-parallel, corrects the output
ANN corrects the by traditional
output model; model;
by traditional (4) the
traditional model output
(4) the traditional is used is
model output upused
to theuptime of time
to the gel effect
of gelappearance, then thethen
effect appearance, output
the
of the ANN model takes over. The latter situation shows the best
output of the ANN model takes over. The latter situation shows the best accuracy of accuracy of prediction,
and the approach is probably transferable to other processes with the change of reaction
mechanism.

5. Processing of the Raw Thermal Analysis Data (Filtering, Classification)

Molecules 2021, 26, 3727 11 of 25

prediction,
Molecules 2021, 26, x FOR PEER REVIEW and the approach is probably transferable to other processes with the change
12 of 26 of
reaction mechanism.

5. Processing of the Raw Thermal Analysis Data (Filtering, Classification)

analyzing the DSC data. The DSC measurements have comprised the cooling stage (from
First, we consider possible applications of ANNs for treatment and analysis of the raw
25 °C to −40 °C at 10 K min−1 rate) followed by the heating step up to 180 °C at 10 K min−1.
thermal analysis data. Sbirrazzuoli et al. [9,16] have proposed using the neural network
Typical DSC curves obtained in this manner are shown in Figure 8. The observed thermal
models for filtering and deconvolution of the DSC data. Caloric thermoanalytical signals
events were analyzed in terms of the onset and peak temperatures and heat effects. Table
have been generated by the Joule effect. The respective electric pulse has been programmed
3 compiles the results for pure milk (the control sample) and milk with different amount
to certain thermodynamic and kinetic parameters. ANN has been trained on the resulting
of the starch additive, the subscripts c, m, and b correspond to the crystallization, melting,
DSC data setting as target values the expected parameters. In this manner, the transfer
and boiling processes.
function of the calorimeter can be obtained and used for deconvolution of DSC data.
As seen from Table 3, the measured quantities are all altered by the additive presence,
Moreover, the results have shown high tolerance of ANNs to a noise introduced in the
and the main purpose of that study is to train the neural network to identify the adulter-
simulated DSC data.
ated sample. This is a chemometric problem of classification, i.e., ANN should output 1
Another important aspect of the ANNs application to the raw TA data is their ability to
for the correct additive type and 0 for another additive types. The ANN (9-3-5) has been
recognize some regularities in the data. Recently, Cruz et al. [40] have proposed to employ
used. The input data are nine measured thermal properties (first column in Table 1), and
ANNs for detecting adulteration in milk. That is, the DSC curves for samples with varying
the outputs are five sample classes (control, starch, formaldehyde, whey, urea). Apart
content of additives (starch, formaldehyde, whey, urea) show certain differences, and the
from ANN, other machine learning methods have been tested, namely, the random forest
general idea is to train the neural network to detect these undesired additives by analyzing
and gradient boosting machine (GBM). Excellent prediction capacity (100%) is accom-◦
the DSC data. The DSC measurements have comprised the cooling stage (from 25 C
plished for ANN and GBM methods. Interestingly, these two top-performing methods
to −40 ◦ C at 10 K min−1 rate) followed by the heating step up to 180 ◦ C at 10 K min−1 .
reveal different values of the input variables importance, e.g., the onset of melting Tom is
Typical DSC curves obtained in this manner are shown in Figure 8. The observed thermal
deemed not important by ANN (only 14%), whereas it is one of most important predictors
events were analyzed in terms of the onset and peak temperatures and heat effects. Table 3
by GBM (84%). Overall, the application of ANN for preliminary analysis of TA results is
compiles the results for pure milk (the control sample) and milk with different amount of
very promising. This example of the classification problem can be easily extended to other
the starch additive, the subscripts c, m, and b correspond to the crystallization, melting,
problems of analytical and industrial relevance.
and boiling processes.

DSC
Figure8.8.DSC
Figure signals
signals forfor milk
milk samples
samples adulterated
adulterated withwith various
various amount
amount of starch
of starch [40]. Thermal
[40]. Thermal
processes
processesthat take
that place
take during
place cooling
during andand
cooling subsequent heating
subsequent are indicated.
heating Reprinted
are indicated. with with
Reprinted
permission
permissionfrom
fromElsevier.
Elsevier.

Table 3.
AsThermal properties
seen from derived
Table 3, from the DSC
the measured data forare
quantities various milk samples
all altered by the [40]. Reprinted
additive presence,
with
andpermission from Elsevier.
the main purpose of that study is to train the neural network to identify the adulterated
sample. This is a Control
Parameters
chemometric problem ofStarch2
Starch1
classification, i.e., ANNStarch4
Starch3
should output 1 for the
Starch5
correct additive type and 0 for another additive types. The ANN (9-3-5) has been used.
Toc (°C) −0.430 0.035 0.065 −0.480 −0.670 −0.535
The input data are nine measured thermal properties (first column in Table 1), and the
Tpc (°C) −0.400 −0.290 −0.035 −0.795 −0.535 −0.740
outputs are five sample classes (control, starch, formaldehyde, whey, urea). Apart from
ΔHc (J g−1) −253.50 −233.430 −247.197 −244.578 −243.339 −244.377
ANN, other machine learning methods have been tested, namely, the random forest and
Tom (°C)
gradient boosting−0.385 −0.645 Excellent
machine (GBM). −0.535 –1.140
prediction capacity –1.925 −1.230
(100%) is accomplished
T pm (°C) 5.315 4.600 5.665 5.495
for ANN and GBM methods. Interestingly, these two top-performing methods 5.350 6.510reveal
ΔH m (J g−1) 220.833 263.704 236.707 199.669 217.466 226.643
different values of the input variables importance, e.g., the onset of melting Tom is deemed
notTimportant
ob (°C) by104.505
ANN (only 103.955 105.645
14%), whereas 105.400
it is one 107.025
of most important 163.480 by
predictors
GBM Tpb(84%).
(°C) Overall,
123.330 118.615 of131.570
the application 123.805
ANN for preliminary 121.935 of TA
analysis 168.880
results is
ΔHb (J g−1) 3577.5 870.990 4313.429 1416.971 1675.230 844.814
Notes: The results show only the mean value, the standard deviation values can be found in the
original paper. Control sample = pure milk without additives, samples called Starch 1, 2, 3, 4, 5 = 1,
Molecules 2021, 26, 3727 12 of 25

very promising. This example of the classification problem can be easily extended to other
problems of analytical and industrial relevance.

Table 3. Thermal properties derived from the DSC data for various milk samples [40]. Reprinted
with permission from Elsevier.

Parameters Control Starch1 Starch2 Starch3 Starch4 Starch5

Toc (◦ C) −0.430 0.035 0.065 −0.480 −0.670 −0.535
Tpc (◦ C) −0.400 −0.290 −0.035 −0.795 −0.535 −0.740
∆Hc (J g−1 ) −253.50 −233.430 −247.197 −244.578 −243.339 −244.377
Tom (◦ C) −0.385 −0.645 −0.535 −1.140 −1.925 −1.230
Tpm (◦ C) 5.315 4.600 5.665 5.495 5.350 6.510
∆Hm (J g−1 ) 220.833 263.704 236.707 199.669 217.466 226.643
Tob (◦ C) 104.505 103.955 105.645 105.400 107.025 163.480
Tpb (◦ C) 123.330 118.615 131.570 123.805 121.935 168.880
∆Hb (J g−1 ) 3577.5 870.990 4313.429 1416.971 1675.230 844.814
Molecules 2021, 26, x FOR PEER REVIEW
Notes: The results show only the mean value, the standard deviation values can be found in the13 of 26 paper.
original
Control sample = pure milk without additives, samples called Starch 1, 2, 3, 4, 5 = 1, 2, 3, 4, 5 g L−1 milk. Subscripts
0 and p denote the onset and peak values, subscripts c, m, and b correspond to crystallization, melting, and
boiling processes.
2, 3, 4, 5 g L−1 milk. Subscripts 0 and p denote the onset and peak values, subscripts c, m, and b
correspond to crystallization, melting, and boiling processes.
Wesolowski et al. [89] have suggested utilizing neural networks for compatibility
assessment. This study appears to be the only example of the Kohonen network [90]
Wesolowski et al. [89] have suggested utilizing neural networks for compatibility as-
(also knownThis
sessment. as self-organizing
study appears tomap, be theSOM) being applied
only example directly to
of the Kohonen the TA[90]
network data. In short,
(also
this
known as self-organizing map, SOM) being applied directly to the TA data. In short, this a two-
type of network has no hidden layer and serves to map the input data into
dimensional
type of network grid.hasSOMs
no hiddendiffer also
layer and inserves
the training
to map the method; they
input data apply
into unsupervised
a two-dimen-
competitive
sional grid. SOMs differ also in the training method; they apply unsupervised competitiveIn their
learning. Figure 9b,c gives examples of the output topological maps.
work [89], the
learning. authors
Figure looked
9b,c gives at the of
examples thermogravimetric
the output topological profiles
maps. ofIncaffeine mixtures
their work [89], with
the authors looked at the thermogravimetric profiles of caffeine
several excipients, e.g., cellulose and glycine. First, the correlation matrix betweenmixtures with several ex- the
cipients, e.g., cellulose and glycine. First, the correlation matrix between
temperatures at certain mass loss values is built. Several temperature values at specific the temperatures
at certain
mass mass loss values
loss percentages (T5 , Tis20built.
, T35 ,Several
T55 , T70temperature
) that showvalues
strongat correlation
specific mass areloss per- and
selected
centages (T 5, T20, T35, T55, T70) that show strong correlation are selected and serve as inputs
serve as inputs in SOM (Figure 9a). The optimal map size was found to be 5 × 5. Figure 9b,c
in SOM (Figure 9a). The optimal map size was found to be 5 × 5. Figure 9b,c shows the
shows the results of calculations for two typical mixtures. The highlighted neuron for the
results of calculations for two typical mixtures. The highlighted neuron for the particular
particular sample is the one that satisfies best the selection criteria. The interpretation of
sample is the one that satisfies best the selection criteria. The interpretation of SOM maps
SOM maps
is based onisthebased on the that
assumption assumption
the samples that theclose
with samples with close
compositions shouldcompositions
nest in alliedshould
nest in allied
regions. Thus, regions. Thus, for caffeine/cellulose,
for caffeine/cellulose, the mixtures 7:3, 9:1 thelie
mixtures
close to 7:3, 9:1 lie
the neat close to the
caffeine.
neat caffeine. Analogously,
Analogously, the system withthe system
excess with excess
of cellulose activatesofthe
cellulose
neuronsactivates the neurons
in the proximity of in
thethat
proximity of that for
for pure cellulose. pure cellulose.
Another Another
mixture, caffeine mixture,
with glycine,caffeine
exhibits with
a much glycine,
different exhibits
a much
behaviordifferent behavior
(Figure 9c), thus, it(Figure
is deemed 9c), thus, it is deemed
incompatible. incompatible.
More mixtures More mixtures
were investigated in
thisinvestigated
were manner, and in thethis
outcome
manner, of SOMand were in agreement
the outcome of SOMwithwere
otherinexperimental
agreementtech- with other
niques (DSC,techniques
experimental FTIR). The (DSC,proposed approach
FTIR). using SOMs
The proposed outperforms
approach using theSOMs twooutperforms
other
thechemometric methods—the cluster
two other chemometric analysis and
methods—the the principal
cluster analysis component analysis.component
and the principal Both
these methods are not capable of resolving the mixtures adequately,
analysis. Both these methods are not capable of resolving the mixtures adequately, when when given the same
characteristic temperatures as input.
given the same characteristic temperatures as input.

Figure
Figure 9. ANN
9. ANN application
application in the
in the compatibility
compatibility assessment
assessment [89]:(a)
[89]: (a)Schematic
Schematicof
of the
the proposed
proposed workflow
workflowand andmapping
mapping with
with ANN of caffeine mixtures with cellulose (b) and glycine (c), mixture ratios are indicated, i.e., 9:1, 7:3, 1:1, 3:7, 1:9.
ANN of caffeine mixtures with cellulose (b) and glycine (c), mixture ratios are indicated, i.e., 9:1, 7:3, 1:1, 3:7, 1:9. Reprinted
Reprinted with permission from Elsevier.
with permission from Elsevier.
6. Thermokinetic Analysis with Neural Networks
In this section, we discuss several applications of ANNs relevant to thermokinetic
analysis. It should be immediately noted that some publications [17,91–93] on using
ANNs for kinetic analysis deal with analysis of single-heating rate data or a single iso-
thermal measurement. This approach is not recommended by the Kinetics Committee of
Molecules 2021, 26, 3727 13 of 25

6. Thermokinetic Analysis with Neural Networks

In this section, we discuss several applications of ANNs relevant to thermokinetic
analysis. It should be immediately noted that some publications [17,91–93] on using ANNs
for kinetic analysis deal with analysis of single-heating rate data or a single isothermal
measurement. This approach is not recommended by the Kinetics Committee of the
International Confederation for Thermal Analysis and Calorimetry (ICTAC) [28], as it does
not allow for properly discriminating between the reaction models and, thus, for obtaining
reliable kinetic parameters, as discussed further. Nevertheless, some of these works are
discussed below to illustrate the development and current status of the ANNs applications
in thermokinetics.
One of the benefits of the ANNs use for kinetic studies [9] is that they could bypass an
explicit usage of Equation (1). This statement needs some clarification. For instance, some
of the examples discussed in Section 3 show that one can predict the conversion degree
as function of temperature and time without invoking Equation (1) and determining the
kinetics triplets at all. However, such predictions are yet to be thoroughly verified to prove
their general validity. On the other hand, evaluation of the kinetic triplets with the aid of
ANN is possible without explicitly separating the k( T ) and f (α) terms as well as making
any assumptions about their mathematical form. Some evaluations of this type [21,94,95]
have already been thoroughly tested in terms of predictions outside of the data space where
the ANN is trained. However, such evaluations are based on the neural networks trained
on the data simulated according to Equation (1), so that the knowledge of Equation (1)
implicitly propagates into predictions.
Gauri et al. [15] compare the performance of the mechanistic kinetic modeling with
the neural network. They consider a heterogeneous reaction of various limestones with
SO2 . Two reaction models have been proposed in literature for this type of reaction: the
shrinking unreacted core model and the distributed pore model. As an example, we will
consider only the simpler one, the shrinking unreacted core model:

dα 3Ks
=  , (6)
dt 1 1 1
+γ −
(1−α)2/3 (1−α)1/3 (1+ Pα)1/3


where Ks = (k s Cs )/( R0 N0 ), γ = ( R0 k s )/De , P = N0 Vp − Vr with ks standing for the
rate constant of surface reaction, N0 —initial concentration of the solid reactant, Cs —the
surface concentration of gas, R0 —the initial radius of the grain, Vp and Vr —the molar
volume of solid product and reactant, De —the diffusivity of Ca2+ ions. The more complex,
distributed, pore model, additionally uses the pore properties that have to be evaluated in
separate structural measurements. Obviously, the application of these mechanistic models
is not straightforward. Moreover, there are some simplifying assumptions used in their
derivation. On the contrary, the neural network model does not need these additional
parameters and relies only on the TA data. Comparison of the fit performance by the two
mechanistic models and by ANN shows that the latter is most accurate.
Another feature of using ANNs is noticed when analyzing a heterogeneous process of
iron oxide reduction [19]. A complicating factor, along with a consecutive chemical reaction
(Fe2 O3 → Fe3 O4 → FeO), is the structural changes of the interface throughout the reaction
(sintering, cracking etc.). As a result, the conversion degree shows non-monotonic behavior,
i.e., the reduction rate is delayed at temperatures about 750 ◦ C (Figure 10). Authors note
that this effect was found by other workers. The benefit of the ANN model is that it
describes all the data, including this effect, whereas the single traditional mechanistic
model fails to do that. Consideration of various reducing gas compositions shows that
ANN also permits fitting the reaction course both in the kinetic and diffusion regimes.
Accomplishing the same with the mechanistic models requires using two different models.
An idea of using the neural networks primarily for determination of the kinetic triplets
has been put forward by Sbirrazzuoli et al. [9] and further developed by Conesa et al. [56],
Muravyev and Pivkina [21], and other more recent workers [61,94–97]. As the approach
 Molecules 2021, 26, 3727 14 of 25

develops, the complexity of the reaction domain has been increasing. The initiatory
work [9] has employed for training and testing the second-order reaction models with
activation energy 74–80 kJ mol−1 , preexponential factor ln(A, s−1 ) = 18–20. Yet, the15recent
Molecules 2021, 26, x FOR PEER REVIEW of 26
study [21] has considered ten reaction models with a large span of the activation energy
and preexponential factor values (Figure 11).

Figure 10. ANN (3-20-1) representation of the conversion degree for Fe2O3 reduction in 90%N2 +
5%CO + 5%H2 gas flow [19]. Reprinted with permission from Elsevier.

An idea of using the neural networks primarily for determination of the kinetic tri-
plets has been put forward by Sbirrazzuoli et al. [9] and further developed by Conesa et
al. [56], Muravyev and Pivkina [21], and other more recent workers [61,94–97]. As the
approach develops, the complexity of the reaction domain has been increasing. The initi-
atory work [9] has employed for training and testing the second-order reaction models
with activation energy 74–80 kJ mol−1, preexponential factor ln(A, s−1) = 18–20. Yet, the
recent study [21] has considered ten reaction models with a large span of the activation
energy and preexponential factor values (Figure 11).
The general idea of the approach is illustrated by Figure 12. The input vector is a set
of sparse TA data (e.g., 20 points [56] or 49 points [21] for each run) at several heating
rates. The output variables are the activation energy Ea, preexponential factor A, and re-
action order n. Conesa et al. [56] also employ ANN to evaluate the kinetic triplets when
using the pyrolysis data of lignin, cellulose, and polyethylene as input. These triplets
when inserted in Equation (1) give rise to the TGA data that match closely the experi-
mental mass
Figure 10. loss
ANN profiles.
(3-20-1) Thus, the neural
representation network here
of the conversion is for
degree utilized
Fe O3 as an alternative
reduction in 90%Nto+
Figure 10. ANN
traditional (3-20-1) optimization
nonlinear of the conversion degree for Fe2O3 2reduction
representation procedure. in 90%N2 + 2
5%CO + 5%H2 gas flow [19]. Reprinted with permission from Elsevier.
5%CO + 5%H2 gas flow [19]. Reprinted with permission from Elsevier.

An idea of using the neural networks primarily for determination of the kinetic tri-
plets has been put forward by Sbirrazzuoli et al. [9] and further developed by Conesa et
al. [56], Muravyev and Pivkina [21], and other more recent workers [61,94–97]. As the
approach develops, the complexity of the reaction domain has been increasing. The initi-
atory work [9] has employed for training and testing the second-order reaction models
with activation energy 74–80 kJ mol−1, preexponential factor ln(A, s−1) = 18–20. Yet, the
recent study [21] has considered ten reaction models with a large span of the activation
energy and preexponential factor values (Figure 11).
The general idea of the approach is illustrated by Figure 12. The input vector is a set
of sparse TA data (e.g., 20 points [56] or 49 points [21] for each run) at several heating
rates. The output variables are the activation energy Ea, preexponential factor A, and re-
action
Figureorder n. Conesa
11. The et al. [56]
plot populated withalso
theemploy ANN tofactor,
(preexponential evaluate the kinetic
activation triplets
energy) whenfor
pairs used
Figure
using 11. The
theand plot
pyrolysispopulated
data with
of (red the
lignin, (preexponential
cellulose, factor, activation
and polyethylene energy) pairs used for
training
training and testingofof
testing ANNs
ANNs (red points),
points), andvalidation
and validation ofthe
of resultsas
theresults input.
(gray
(gray and
These
andblue triplets
bluepoints)
points) [21].
[21].
when inserted
Reprinted with in Equationfrom
permission (1) Elsevier.
give rise to the TGA data that match closely the experi-
Reprinted with permission from Elsevier.
mental mass loss profiles. Thus, the neural network here is utilized as an alternative to
traditional nonlinear
The general optimization
idea procedure.
of the approach is illustrated by Figure 12. The input vector is a set
of sparse TA data (e.g., 20 points [56] or 49 points [21] for each run) at several heating rates.
The output variables are the activation energy Ea , preexponential factor A, and reaction
order n. Conesa et al. [56] also employ ANN to evaluate the kinetic triplets when using the
pyrolysis data of lignin, cellulose, and polyethylene as input. These triplets when inserted
in Equation (1) give rise to the TGA data that match closely the experimental mass loss
profiles. Thus, the neural network here is utilized as an alternative to traditional nonlinear
optimization procedure.
Aghbashlo et al. [96,97] adopt the same approach but utilize a more advanced math-
ematical apparatus. The authors first use the principal component analysis to reduce
the number of input variables from six (C, H, N, O, S content, and heating rate) to three.
Then these transformed input data are fed into an adaptive neuro-fuzzy inference sys-
tem (ANFIS) to produce the output kinetic parameters Ea , A, n. ANFIS [98] represents a

Figure 11. The plot populated with the (preexponential factor, activation energy) pairs used for
training and testing of ANNs (red points), and validation of the results (gray and blue points) [21].
Reprinted with permission from Elsevier.
Molecules 2021, 26, 3727 15 of 25

combination of neural network model with a fuzzy logic (FL) system. Additionally, the
authors use the genetic algorithm (GA) to optimize the ANFIS parameters. The developed
model results in the kinetic parameters that agree well with those obtained by traditional
optimization procedure for five various biomass feedstocks (bamboo, Phyllanthus emblica
Molecules 2021, 26, x FOR PEER REVIEW 16 of 26
kernel, Musa balbisiana, cattle manure, peanut shell). Although the computational approach
from the study (ANN+FL+GA) is rather advanced, the major limitation is the constrained
training that is limited to single heating rate and reaction order model.

Figure 12. The framework of thermokinetic analysis with ANN proposed in Ref. [56]. Reprinted with permission
Figure 12. The framework of thermokinetic analysis with ANN proposed in Ref. [56]. Reprinted with permission from
from Elsevier.
Elsevier.
Muravyev and Pivkina [21] extend the above-described approach [56] by considering
Aghbashlo
a broad range ofetthe al. [96,97]
activation adopt the same
energy and approach but utilize
preexponential factor a more
pairs advanced
(Figure 11). math-
Ad-
ematical
ditionally,apparatus.
instead ofThe n-th authors first use model,
order reaction the principal
10 othercomponent
reaction modelsanalysisare to reduce
taken into the
number
account.ofTable
input4 variables
summarizes fromthe sixresults
(C, H, ofN,their
O, S content,
study. Itand heating
is found thatrate)
the to three.
best Then
accuracy
these transformed
(determined as ANN input data are fedout
performance into
ofan theadaptive
domain neuro-fuzzy
space used for inference
trainingsystem (AN-
and testing,
FIS) to produce
Figure the output
11) is achieved whenkinetic parameters
the specific ANN isEatrained
, A, n. ANFIS [98] represents
to determine only one acomponent
combina-
tion of kinetic
of the neural triplet
network (Eamodel
, A, f (αwith a fuzzy logicit(FL)
)). Additionally, system.
is shown thatAdditionally,
the suggested theapproach
authors
use the genetic
confirms some algorithm
previously(GA) to optimize
established facts:the(1)ANFIS
failureparameters.
of the kinetic The developed
analysis based model
on a
single heating
results rate data
in the kinetic [28]; (2) enhanced
parameters that agree capability
well withof controlled
those obtained rate thermal analysisopti-
by traditional [99]
to discriminate
mization procedure among the reaction
for five models. feedstocks
various biomass Conceptually, since Phyllanthus
(bamboo, this approach involves
emblica ker-
the synthetic
nel, data ascattle
Musa balbisiana, input,manure,
it pavespeanut
the way to theAlthough
shell). reinforcement learning instead
the computational of the
approach
supervised
from the studylearning for neural isnetwork.
(ANN+FL+GA) That is, one
rather advanced, theroutine generatesisthermal
major limitation analysis
the constrained
profiles for
training thatknown
is limitedkinetic parameters,
to single heating while
rate another (ANN)
and reaction usesmodel.
order this large amount of data
for training.
Muravyev and Pivkina [21] extend the above-described approach [56] by considering
a broadHuang
rangeet of
al. the
[94]activation
have adopted energythe and
samepreexponential
approach, but further increased
factor pairs (Figurethe 11).
number Addi- of
reaction types
tionally, insteadto of
17,n-th
and order
utilized a moremodel,
reaction advanced mathematical
10 other reactiontool, models the general
are taken regression
into ac-
neural Table
count. network. Most importantly,
4 summarizes the authors
the results of their employ
study.datasets
It is foundof simulated
that the nonisothermal
best accuracy
(determined as ANN performance out of the domain space used for trainingheating
experiments at several different heating rates. They stress that only at three rates
and testing,
one can achieve an acceptable recognition accuracy for the
Figure 11) is achieved when the specific ANN is trained to determine only one component reaction model and that the
accuracy improves significantly when the number of experiments
of the kinetic triplet (Ea, A, 𝑓(𝛼)). Additionally, it is shown that the suggested approach (heating rates) used in
ANN training is five. The relative error in the output preexponential
confirms some previously established facts: (1) failure of the kinetic analysis based on a factor and activation
energyheating
single values rateis asdata
low [28];
as 4%(2) forenhanced
the simulated data.ofAdditionally,
capability controlled rate thethermal
authors analysis
applied
the developed neural network to the experimental data for
[99] to discriminate among the reaction models. Conceptually, since this approach in-Li 4 Ti 5 O 12 /C high-temperature
reaction.
volves theThe ANN output
synthetic data asagrees
input,well with the
it paves the literature
way to the results [100] of these
reinforcement data analysis
learning instead
accomplished with isoconversional and model-fitting kinetic
of the supervised learning for neural network. That is, one routine generates thermal approaches. It is noteworthyanal-
that profiles
ysis according fortoknown
the isoconversional
kinetic parameters, analysis, the activation
while another (ANN) energyuses of the
thisprocess is more
large amount
or less constant throughout the reaction, thus, indicating the single-step reaction kinetics.
of data for training.
Molecules 2021, 26, 3727 16 of 25

Table 4. ANN details and performance of models developed for kinetic analysis in Ref. [21].
Reprinted with permission from Elsevier.

Input Output P, %
# Cases Nh AC, %
Parameters Ni Parameters No Training Test
T(α)
1 49 Reaction model 6 3278 15 99.1 98.2 88
B = 5 K/min
T(α)
2 49 Ea 1 6106 32 93.8 92.2 62
B = 5 K/min
3 T(α) β = 0.5, 20 K/min 98 Reaction model 6 199 18 99.8 99.4 100
4 T(α) β = 0.5, 20 K/min 98 Ea 1 147 14 99.9 99.9 98
5 T(α) β = 0.5, 20 K/min 98 lgA 1 139 10 99.9 99.9 98
98(E),
6 T(α) β = 0.5, 20 K/min 98 Ea , lgA 2 147 13 99.9 99.9
92(A)
71 T(α) β = 0.5, 20 K/min 50 Reaction model 10 1109 20 99.8 99.4 100
T(α) β = 0.5, 5,
81 75 Reaction model 10 483 25 99.8 99.1 100
20 K/min
9 T(α) c = 1·10−3 min−1 49 Reaction model 6 851 27 90.1 84.3 65
10 2 T(α) c = 1·10−3 min−1 49 Reaction model 6 678 15 99.9 99.9 100
81 (E),
Reaction model,
11 2 T(α) c = 1·10−3 min−1 49 8 678 21 99.6 99.2 44(A),
Ea , lgA
100 (R)
Notes: Ni , Nh and No are the number of neurons in input, hidden, and output layer of network, P and AC—the
performance and accuracy of the respective ANN. 1 For these models, the reaction types (B1, D2, D3, F2, L2, F1,
P2, R3, A2, A3) were used, and the step for conversion degree ∆α = 0.05 was applied instead of ∆α = 0.02. 2 For
these models, the reaction types (F1, A2, A3, D1, D3, B1) were used instead of (F1, R2, A3, F2, D3, B1).

Kuang and Xu [95] have implemented the most advanced mathematical approach
to this problem, i.e., the one-dimensional convolutional neural network (CNN). As in
the previously considered work [94], the single neural network was utilized to estimate
the whole kinetic triplet. The application of the trained CNN model for estimating the
preexponential factor and activation energy has resulted in less than 3% error in case of
using a dataset of three heating rates for training. Remarkably, the authors have used
noisy input data to compare the CNN model with the one used in a previous study, the
feedforward neural network (more specifically, multi linear perceptron, MLP). Figure 13
displays that with increasing the amplitude of the Gaussian noise, the performance of MLP
starts decreasing, which does not apply to the CNN model. Obviously, the convolutional
neural network shows superior robustness toward the noisy data. Finally, the model
has been applied to the pine wood pyrolysis data. The details of calculations (e.g., the
isoconversional plots) are not provided, but apparently all the discussed kinetic data are
associated with the first mass loss stage (corresponds to ~85% of total mass loss). The kinetic
parameters obtained with the CNN agree well with those computed by two isoconversional
methods, Kissinger–Akahira–Sunose and Flynn–Wall–Ozawa methods.
It should be mentioned that ANN does not always yield the results that agree well
with the isoconversional ones. These two approaches have been compared as applied
to the kinetics in the system lumefantrine/molecularly imprinted polymer [101]. The
isoconversional method has demonstrated that in the range of conversions from 0.05 to
0.90, the activation energy rises from roughly 80 to 150 kJ mol−1 . This is a sure sign of
the kinetics being complex, i.e., to involve at least two steps with significantly different
activation energies [102]. Yet, the ANN estimates yield the activation energy whose values
fluctuate in the range of approximately 160–200 kJ mol−1 depending on the heating rate
used in analysis [101]. Of course, one cannot expect ANNs to reveal the multi-step nature
of the process kinetic as long as the networks are trained on single-step kinetics. Likewise,
the usage of single heating rate ANN analysis and training is a faulty approach to kinetic
calculations as already demonstrated in a previous study [28].
Molecules 2021, 26, x FOR PEER REVIEW 18 of 26

with the first mass loss stage (corresponds to ~85% of total mass loss). The kinetic param-
Molecules 2021, 26, 3727 17 of 25
eters obtained with the CNN agree well with those computed by two isoconversional
methods, Kissinger–Akahira–Sunose and Flynn–Wall–Ozawa methods.

Figure 13. Comparison of the accuracy of convolutional neural network (CNN) and multilinear
Figure 13. Comparison of the accuracy of convolutional neural network (CNN) and multilinear
perceptron (MLP) models
perceptron (MLP) models asas aa function
function of
of the
the introduced
introducednoise
noisein
inthe
theinput
inputdata
data[95].
[95].Reprinted
Reprintedwith
permission from Elsevier.
with permission from Elsevier.

A principally different attempt to perform kinetic analysis with ANNs has been made
It should be mentioned that ANN does not always yield the results that agree well
in a series of papers by Sebastiao and coworkers. The general idea is to utilize the neurons
with the isoconversional ones. These two approaches have been compared as applied to
in the form of reaction model functions. The authors apply a similar approach for the
the kinetics in the system lumefantrine/molecularly imprinted polymer [101]. The isocon-
description of isothermal processes of widely differing materials, i.e., rhodium (II) ac-
versional method has demonstrated that in the range of conversions from 0.05 to 0.90, the
etate [17,18], efavirenz and lamivudine [93], thalidomide [103], polyurethane [104], hybrid
activation energy rises from roughly 80 to 150 kJ mol−1. This is a sure sign of the kinetics
of carbon nanotubes with poly(3-hexylthiophene) [105], composite Li4 SiO4 /MgO [106],
being complex, i.e., to involve at least two steps with significantly different activation en-
calcium trihydrate/furosemide system [107], and lumefantrine/molecularly imprinted
ergies [102]. Yet, the ANN estimates yield the activation energy whose values fluctuate in
polymer [101]. To show the basics of this method, we will follow the first paper [18],
the range of approximately 160–200 kJ mol−1 depending on the heating rate used in anal-
where the ANN with a single neuron in the hidden layer is considered (Figure 14). The
ysis [101]. Of course, one cannot expect ANNs to reveal the multi-step nature of the pro-
neural network is fed with the time data from an isothermal measurement and outputs
cess kinetic as long
the conversion as the
degree networks
value. Givenare thetrained
weightsonand single-step
biases, askinetics.
shown inLikewise,
Figure 13, theand
usage
the
of single heating rate ANN analysis and training
sigmoidal activation function, the output value is: is a faulty approach to kinetic calcula-
tions as already demonstrated in a previous study [28].
A principally different attempt w32 kinetic analysis with ANNs has been made
o = to perform
−(
+ w30 . (7)
+ e w21 i+wThe
in a series of papers by Sebastiao and1coworkers. 20 ) general idea is to utilize the neurons

in the form of reaction model functions. The authors apply a similar approach for the de-
Equation (7) can be compared with the best mechanistic kinetic model of Prout and
scription of isothermal processes of widely differing materials, i.e., rhodium (II) acetate
Tompkins [108]:
[17,18], efavirenz and lamivudine [93], thalidomide [103], polyurethane [104], hybrid of
carbon nanotubes dαwith poly(3-hexylthiophene)
α [105], composite Li 1 4SiO4/MgO [106], cal-
= kα(1 −system
cium trihydrate/furosemide α) → ln [107], = kt + C → α = . (8)
dt 1 − αand lumefantrine/molecularly
1 + e kt+C) imprinted poly-
−(
mer [101]. To show the basics of this method, we will follow the first paper [18], where the
ANNUsingwith athe similarity
single neuronof in
thethe
Equations
hidden (7) andis(8),
layer the logarithm
considered of the
(Figure 14).optimized
The neural weight
net-
values
work is w21 for
fed witheach
therun
timeis data
plotted
fromagainst the reciprocal
an isothermal temperature
measurement andto give the
outputs theactivation
conver-
energy
sion value.
degree A larger
value. Given number of the and
the weights adjustable
biases,parameters
as shown ininFigureEquation (7) as
13, and compared
the sigmoidal to
the traditional Equation (8) results
activation function, the output value is: in several times smaller residual error for the model.
The idea of using the activation functions 𝑤32
to mimic the mechanistic reaction models is
further advanced in another study𝑜by = Sebastiao
1+𝑒 −(𝑤21 𝑖+𝑤 ) + [17].
et20al. 𝑤30 . More neurons have been added (7)
to the hidden layer, but their activation functions are set in an unusual, for ANNs, way.
Equation (7) can be compared with the best mechanistic kinetic model of Prout and
These activation functions have been selected to represent various kinetic models, e.g.:
Tompkins [108]:
𝑑𝛼 𝛼 1/m 1
ln(→1 −
[−𝛼)
= 𝑘𝛼(1 − 𝑙𝑛 α)] = 𝑘𝑡=+kt𝐶+→C,𝛼 = . (9)
(8)
𝑑𝑡 1-𝛼 1+𝑒 −(𝑘𝑡+𝐶)

for Kolmogorov–Johnson–Mehl–Avrami–Erofeev nucleation-growth reaction [108], or:

2α 2
1− − (1 − α) 3 = kt + C, (10)
3
 e.g.,:
[−ln⁡(1 − 𝛼)]1/𝑚 = 𝑘𝑡 + 𝐶, (9)
for Kolmogorov–Johnson–Mehl–Avrami–Erofeev nucleation-growth reaction [108], or:
2
Molecules 2021, 26, 3727 2𝛼 18 of 25
1− − (1 − 𝛼)3 = 𝑘𝑡 + 𝐶, (10)
3

for Ginstling–Brounshtein three-dimensional diffusion model [108]. That is, first the ideal
reaction model (e.g., one of Equations (8)–(10)) is optimized on the experimental data,
for Ginstling–Brounshtein three-dimensional diffusion model [108]. That is, first the ideal
then the resulting parameters (k, C) are set as weight and bias for the respective neuron
reaction model (e.g., one of Equations (8)–(10)) is optimized on the experimental data, then
(e.g., as w21 and w20 for neuron designated as “2” in Figure 14a). Then, instead of the usual
the resulting parameters (k, C) are set as weight and bias for the respective neuron (e.g., as
training of ANN using the back-propagation algorithm, the Levenberg–Marquardt opti-
w21 and w20 for neuron designated as “2” in Figure 14a). Then, instead of the usual training
mization is performed with weights and biases for the hidden layer (i.e., w21 and w20) fixed,
of ANN using the back-propagation algorithm, the Levenberg–Marquardt optimization
but adjusting the interconnection weights in the output layer (e.g., w32 in Figure 14a, note,
is performed with weights and biases for the hidden layer (i.e., w21 and w20 ) fixed, but
that here no biases like w30 are considered for the outer neuron). From these weights the
adjusting the interconnection weights in the output layer (e.g., w32 in Figure 14a, note, that
relative
here nocontribution
biases like wof the individual reaction models to the whole process can be ob-
30 are considered for the outer neuron). From these weights the relative
tained. By way of example,
contribution of the individual Figure 14b models
reaction represents such
to the contributions
whole process canfor
be several isother-
obtained. By way
mal experiments on the thermal decomposition of lamivudine [93]. Here, the
of example, Figure 14b represents such contributions for several isothermal experimentsthree-dimen-
sional
on thediffusion
thermalmodel (Equation of
decomposition (8))lamivudine
provides the main
[93]. contribution
Here, to the process.diffusion
the three-dimensional Over-
all,model
the above discussed approach can be called as kinetic deconvolution by
(Equation (8)) provides the main contribution to the process. Overall, the aboveneural net-
work.
discussed approach can be called as kinetic deconvolution by neural network.

Figure
Figure 14.14. Kinetic
Kinetic analysis
analysis with
with ANNs
ANNs comprisingthe
comprising theactivation
activationfunctions
functionsfor
forhidden
hiddenlayer
layerneu-
neurons
rons having the mathematical form of ideal reaction types. (a) The scheme of simple ANN modelused
having the mathematical form of ideal reaction types. (a) The scheme of simple ANN model
in Ref.
used [18].
in Ref. (b) (b)
[18]. Relative contributions
Relative of kinetic
contributions models
of kinetic modelsin in
thethe
course of of
course thermal decomposition
thermal decomposi- of
tion of lamivudine, squares depict three-dimensional diffusion model, crosses—linear
lamivudine, squares depict three-dimensional diffusion model, crosses—linear contraction contraction
model,
circles—second-order KJMAE nucleation-growth, triangles—Prout–Tompkins, asterisk—first-order
reaction [93]. Reprinted by permission from Springer Nature.

As seen from above, the kinetic applications of ANN still are in their early stages of
development. There remains much to be done for these applications to match the capa-
bilities of the well-established isoconversional and model-fitting techniques [28,102]. In
regard to this, there are two important principles that must be followed by any computa-
tional technique so that it can produce reliable kinetic triplets. First, in its computations,
a technique must use simultaneously multiple heating rates or, more generally, multiple
temperature programs [28]. For kinetic applications of ANN, it means that the network
training sets must include data obtained at more than one heating rate. A failure to pro-
duce reliable kinetic triplets has already been demonstrated [21] for ANN trained on the
data sets obtained at single heating rates. This problem is yet another manifestation of
the fundamental failure of the kinetic techniques based on single heating rate [109]. In
the case of the traditional isoconversional and model-fitting techniques, reliable kinetic
evaluations require one to use simultaneously no less than 3–4 heating rates. The same
principle should be adhered to when training and using ANNs for kinetic applications.
Molecules 2021, 26, 3727 19 of 25

Second, a computational technique must be capable of detecting and treating multi-step

kinetics [102]. The reality of thermally stimulated processes in the condensed phase is such
that they typically include more than one kinetic step. However, the literature analysis
indicates that, so far, the ANNs used for kinetic purposes have been trained exclusively
on single-step kinetics. Such restricted training limits drastically the application area of
ANNs. To remove this important limitation, proper strategies for multi-step kinetics train-
ing ought to be developed. In the meantime, it seems prudent to support the usage of the
ANNs trained on single-step kinetics by verifying whether the kinetics studied is actually
a single-step one. The latter is readily confirmed as the absence of any significant variation
in the isoconversional activation energy [102].

7. Other Applications Relevant to Thermal Analysis Studies

This section gives a short overview of papers that do not directly focus on thermal
analysis data or thermokinetic parameters, but that can potentially find some applications
in this field. First, neural networks are used to interpolate and extrapolate the tempera-
ture dependency of desired property. Thus, Wunderlich et al. [13] have proposed using
ANNs for prediction of the heat capacity (cp ) data. Note that this property is normally
measured with the thermal analysis method of DSC, and it is of importance to interpolate
and extrapolate the results to particular values of temperature. The experimental heat
capacity values for 15 common polymers have been taken at every 10 K over the 110–360 K
temperature range. These values serve as inputs for neural network (25-15-10), and the
output was cp at T = 10, 20, . . . , 100 K. The reported errors of trained ANN in the cp data
for five polymers from test set are less than 0.8%. Thus, the accuracy of this extrapolation
procedure is better than typical experimental errors (1–5%). Another group of applications
comprises predictions of thermal properties with ANNs. This topic is vast and is not
considered here in detail. It worth mentioning, though, that artificial intelligence methods
have been applied for prediction of thermal conductivity [110,111], vapor pressure [112],
flammability [113], density [114], thermal stability [115], and mechanical properties [116].
Third group of applications includes some developments in experimental devices that can
be potentially applied in future thermal analyzers and relevant instruments. For instance,
Gao et al. [117] discuss an optimal iterative controller for nonlinear processes based on
ANN model (note, that this topic has been widely studied, e.g., [118,119]). The authors
demonstrate that the proposed system offers a better temperature control as compared to a
standard PID controller. This method could be used in thermal analysis equipment where
the temperature control is critical, e.g., in accelerating rate calorimeters [120].

8. Conclusions
Artificial neural networks (ANNs) are a system of powerful tools for solving data
analysis problems in a broad variety of applications. Regarding the applications considered
in the present review, they appear to be in an initiatory stage and use primarily the feed-
forward neural networks as a computational tool. Beyond ANNs, some other artificial
intelligence approaches have been sporadically applied to the TA problems and are worth
mentioning. They include the expert analysis of thermogravimetric data [121], genetic
approach [122] for determining kinetic parameters, modeling with adaptive neuro-fuzzy
inference system [96,123,124] to predict the mass loss data, extreme gradient boosting
algorithm [125] for product yield evaluation. A wider application of machine learning in
these areas is expected as the computational tools become more readily available.
Our review indicates that the most popular ANN applications are associated with
predicting pyrolysis product conversion. This popularity is driven by an obvious practical
importance of the problem. Its solution requires carefully accounting for the TA data
complexity via employing ANNs of an increasingly complex topology. In general, the
performance of ANN models is quite good when assessed within the domain close to that
used for the model development. However, evaluation of the performance outside such
Molecules 2021, 26, 3727 20 of 25

domain is almost absent in the literature. Studies on the extrapolation capabilities of the
ANN models should certainly be encouraged.
Neural network models can also be applied to analyze the TA data directly as shown in
examples of utilizing ANNs for identifying milk samples and probing compatibility of drug
excipients. These applications can be readily extended to a large variety of multi-component
materials, whose characteristic TA parameters change discretely with composition. Such
applications are of obvious relevance to composition optimization and quality control. This
type of usage of ANNs in assaying the TA data is definitely worthy of further exploration,
especially in combination with some approaches from chemometrics, spectroscopy, and
other fields such as filtering, baseline correction, deconvolution, classification, and so on.
In the area of the kinetic applications of ANNs, one should clearly differentiate
between two groups of studies: (i) those that use single heating rate or single isothermal run
for deriving the Arrhenius parameters; and (ii) those that derive the Arrhenius parameters
via simultaneously using multiple heating rates or multiple temperature programs in
general. The results of the first group must be treated with great caution because it is now
established that single temperature program data are fundamentally incapable of yielding
reliable kinetic parameters regardless of the computational technique used. A failure of
such approach has already been demonstrated by comparing ANNs employing single and
multiple heating rates. Only the multiple heating rates approach has proved to be correct,
and this must be kept in mind while developing further kinetic applications of ANNs.
Another crucial trait that must be implemented in future developments is the ability to
recognize and treat the multi-step kinetics.
Overall, the ANN applications present a growing trend in the area covered by this
review. ANN’s provide newer more sophisticated and flexible mathematical tools that
are poised to accomplish a greater level of detail in the process description. Nevertheless,
a greater level of detail does not necessarily translate to qualitatively deeper levels of
understanding. Indeed, the latter are the true milestones of scientific progress. Only time
will tell if ANNs spurred any considerable progress in the field under review. On the other
hand, one must realize that no progress can ever be accomplished without trying new
approaches and tools.

Author Contributions: Conceptualization, N.V.M., H.L.O.J., R.S. and G.L.; investigation, H.L.O.J.,
R.S. and G.L.—original draft preparation, N.V.M.; writing—review and editing, S.V. All authors have
read and agreed to the published version of the manuscript.
Funding: N.V.M. acknowledges the core funding from the Russian Federal Ministry of Science and
Higher Education (0082-2018-0002, AAAA-A18-118031490034-6).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Wilbraham, L.; Mehr, S.H.M.; Cronin, L. Digitizing Chemistry Using the Chemical Processing Unit: From Synthesis to Discovery.
Acc. Chem. Res. 2021, 54, 253–262. [CrossRef]
2. Walters, W.P.; Barzilay, R. Applications of Deep Learning in Molecule Generation and Molecular Property Prediction. Acc. Chem.
Res. 2021, 54, 263–270. [CrossRef] [PubMed]
3. Tkatchenko, A. Machine learning for chemical discovery. Nat. Commun. 2020, 11, 4125. [CrossRef]
4. Yamada, H.; Liu, C.; Wu, S.; Koyama, Y.; Ju, S.; Shiomi, J.; Morikawa, J.; Yoshida, R. Predicting Materials Properties with Little
Data Using Shotgun Transfer Learning. ACS Cent. Sci. 2019, 5, 1717–1730. [CrossRef] [PubMed]
5. Sieniutycz, S.; Szwast, Z. Neural Networks—A Review of Applications. In Optimizing Thermal, Chemical, and Environmental
Systems; Elsevier: Amsterdam, The Netherlands, 2018; pp. 109–120; ISBN 978-0-12-813582-2.
6. Xu, A.; Chang, H.; Xu, Y.; Li, R.; Li, X.; Zhao, Y. Applying artificial neural networks (ANNs) to solve solid waste-related issues: A
critical review. Waste Manag. 2021, 124, 385–402. [CrossRef]
7. McCulloch, W.S.; Pitts, W. A logical calculus of the ideas immanent in nervous activity. Bull. Math. Biol. 1943, 5, 115–133. [CrossRef]
Molecules 2021, 26, 3727 21 of 25

8. Cybenko, G. Approximation by superpositions of a sigmoidal function. Math. Control. Signals Syst. 1989, 2, 303–314. [CrossRef]
9. Sbirrazzuoli, N.; Brunel, D. Computational neural networks for mapping calorimetric data: Application of feed-forward neural
networks to kinetic parameters determination and signals filtering. Neural Comput. Appl. 1997, 5, 20–32. [CrossRef]
10. Egmont-Petersen, M.; de Ridder, D.; Handels, H. Image processing with neural networks—A review. Pattern Recognit. 2002, 35,
2279–2301. [CrossRef]
11. Cammarata, L.; Fichera, A.; Pagano, A. Neural prediction of combustion instability. Appl. Energy 2002, 72, 513–528. [CrossRef]
12. Cho, S.; No, K.; Goh, E.; Kim, J.; Shin, J.; Joo, Y.; Seong, S. Optimization of Neural Networks Architecture for Impact Sensitivity of
Energetic Molecules. Bull. Korean Chem. Soc. 2005, 26, 399–408. [CrossRef]
13. Darsey, J.A.; Noid, D.W.; Wunderlich, B.; Tsoukalas, L. Neural-net extrapolations of heat capacities of polymers to low tempera-
tures. Die Makromol. Chem. Rapid Commun. 1991, 12, 325–330. [CrossRef]
14. Ventura, S.; Silva, M.; Perez-Bendito, D.; Hervas, C. Multicomponent Kinetic Determinations Using Artificial Neural Networks.
Anal. Chem. 1995, 67, 4458–4461. [CrossRef] [PubMed]
15. Bandyopadhyay, J.K.; Annamalai, S.; Gauri, K.L. Application of artificial neural networks in modeling limestone–SO2 reaction.
AIChE J. 1996, 42, 2295–2302. [CrossRef]
16. Sbirrazzuoli, N.; Brunel, D.; Elegant, L. Neural networks for kinetic parameters determination, signal filtering and deconvolution
in thermal analysis. J. Therm. Anal. Calorim. 1997, 49, 1553–1564. [CrossRef]
17. Sebastião, R.C.O.; Braga, J.P.; Yoshida, M.I. Competition between kinetic models in thermal decomposition: Analysis by artificial
neural network. Thermochim. Acta 2004, 412, 107–111. [CrossRef]
18. Sebastiao, R.C.O.; Braga, J.P.; Yoshida, M.I. Artificial neural network applied to solid state thermal decomposition. J. Therm. Anal.
Calorim. 2003, 74, 811–818. [CrossRef]
19. Wiltowski, T.; Piotrowski, K.; Lorethova, H.; Stonawski, L.; Mondal, K.; Lalvani, S.B. Neural network approximation of iron oxide
reduction process. Chem. Eng. Process. Process Intensif. 2005, 44, 775–783. [CrossRef]
20. Bezerra, E.M.; Bento, M.S.; Rocco, J.A.F.F.; Iha, K.; Lourenço, V.L.; Pardini, L.C. Artificial neural network (ANN) prediction of
kinetic parameters of (CRFC) composites. Comput. Mater. Sci. 2008, 44, 656–663. [CrossRef]
21. Muravyev, N.V.; Pivkina, A.N. New concept of thermokinetic analysis with artificial neural networks. Thermochim. Acta 2016, 637,
69–73. [CrossRef]
22. Cavalheiro, É.T.G. Thermal Analysis. In Reference Module in Chemistry, Molecular Sciences and Chemical Engineering; Elsevier:
Amsterdam, The Netherlands, 2018, ISBN 978-0-12-409547-2.
23. Sun, Y.; Liu, L.; Wang, Q.; Yang, X.; Tu, X. Pyrolysis products from industrial waste biomass based on a neural network model.
J. Anal. Appl. Pyrolysis 2016, 120, 94–102. [CrossRef]
24. Hough, B.R.; Beck, D.A.; Schwartz, D.T.; Pfaendtner, J. Application of machine learning to pyrolysis reaction networks: Reducing
model solution time to enable process optimization. Comput. Chem. Eng. 2017, 104, 56–63. [CrossRef]
25. Hua, F.; Fang, Z.; Qiu, T. Application of convolutional neural networks to large-scale naphtha pyrolysis kinetic modeling. Chin. J.
Chem. Eng. 2018, 26, 2562–2572. [CrossRef]
26. Šesták, J. Ignoring heat inertia impairs accuracy of determination of activation energy in thermal analysis. Int. J. Chem. Kinet.
2019, 51, 74–80. [CrossRef]
27. Vyazovkin, S. How much is the accuracy of activation energy affected by ignoring thermal inertia? Int. J. Chem. Kinet. 2020, 52,
23–28. [CrossRef]
28. Vyazovkin, S.; Burnham, A.K.; Criado, J.M.; Perez-Maqueda, L.A.; Popescu, C.; Sbirrazzuoli, N. ICTAC Kinetics Committee
recommendations for performing kinetic computations on thermal analysis data. Thermochim. Acta 2011, 520, 1–19. [CrossRef]
29. Burnham, A.K.; Dinh, L.N. A comparison of isoconversional and model-fitting approaches to kinetic parameter estimation and
application predictions. J. Therm. Anal. Calorim. 2007, 89, 479–490. [CrossRef]
30. Muravyev, N.V.; Melnikov, I.N.; Monogarov, K.A.; Kuchurov, I.V.; Pivkina, A.N. The power of model-fitting kinetic analysis
applied to complex thermal decomposition of explosives: Reconciling the kinetics of bicyclo-HMX thermolysis in solid state and
solution. J. Therm. Anal. Calorim. 2021. [CrossRef]
31. Vyazovkin, S. Modern Isoconversional Kinetics: From Misconceptions to Advances. In Handbook of Thermal Analysis and
Calorimetry; Elsevier: Amsterdam, The Netherlands, 2018; Volume 6, pp. 131–172; ISBN 978-0-444-64062-8.
32. Sbirrazzuoli, N. Model-free isothermal and nonisothermal predictions using advanced isoconversional methods. Thermochim.
Acta 2021, 697, 178855. [CrossRef]
33. Kiselev, V.G.; Muravyev, N.V.; Monogarov, K.A.; Gribanov, P.S.; Asachenko, A.F.; Fomenkov, I.V.; Goldsmith, C.F.; Pivkina, A.N.;
Gritsan, N.P.; Asachenko, A.F. Toward reliable characterization of energetic materials: Interplay of theory and thermal analysis in
the study of the thermal stability of tetranitroacetimidic acid (TNAA). Phys. Chem. Chem. Phys. 2018, 20, 29285–29298. [CrossRef]
34. Koga, N. Physico-Geometric Approach to the Kinetics of Overlapping Solid-State Reactions. In Handbook of Thermal Analysis and
Calorimetry; Elsevier: Amsterdam, The Netherlands, 2018; Volume 6, pp. 213–251; ISBN 978-0-444-64062-8.
35. Opfermann, J.; Hädrich, W. Prediction of the thermal response of hazardous materials during storage using an improved
technique. Thermochim. Acta 1995, 263, 29–50. [CrossRef]
36. Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat.
Mach. Intell. 2019, 1, 206–215. [CrossRef]
Molecules 2021, 26, 3727 22 of 25

37. Novak, R.; Bahri, Y.; Abolafia, D.A.; Pennington, J.; Sohl-Dickstein, J. Sensitivity and Generalization in Neural Networks: An
Empirical Study. arXiv 2018, arXiv:1802.08760.
38. Psichogios, D.C.; Ungar, L.H. A hybrid neural network-first principles approach to process modeling. AIChE J. 1992, 38,
1499–1511. [CrossRef]
39. Wolpert, D.H. The Lack of a Priori Distinctions between Learning Algorithms. Neural Comput. 1996, 8, 1341–1390. [CrossRef]
40. Farah, J.S.; Cavalcanti, R.N.; Guimarães, J.T.; Balthazar, C.F.; Coimbra, P.T.; Pimentel, T.C.; Esmerino, E.A.; Duarte, M.C.K.;
Freitas, M.Q.; Granato, D.; et al. Differential scanning calorimetry coupled with machine learning technique: An effective
approach to determine the milk authenticity. Food Control 2021, 121, 107585. [CrossRef]
41. Ali, J.M.; Hussain, M.; Tade, M.O.; Zhang, J. Artificial Intelligence techniques applied as estimator in chemical process—A
literature survey. Expert Syst. Appl. 2015, 42, 5915–5931. [CrossRef]
42. Da Silva, I.N.; Hernane Spatti, D.; Andrade Flauzino, R.; Liboni, L.H.B.; dos Reis Alves, S.F. Artificial Neural Networks. A Practical
Course; Springer: Cham, Switzerland, 2017; ISBN 978-3-319-43161-1.
43. Graupe, D. Principles of Artificial Neural Networks: Basic Designs to Deep Learning; World Scientific: Singapore, 2019; ISBN
9789811201226.
44. Çepelioğullar, Ö.; Mutlu, I.; Yaman, S.; Haykiri-Acma, H. A study to predict pyrolytic behaviors of refuse-derived fuel (RDF):
Artificial neural network application. J. Anal. Appl. Pyrolysis 2016, 122, 84–94. [CrossRef]
45. Naqvi, S.R.; Hameed, Z.; Tariq, R.; Taqvi, S.A.; Ali, I.; Niazi, M.B.; Noor, T.; Hussain, A.; Iqbal, N.; Shahbaz, M. Synergistic effect
on co-pyrolysis of rice husk and sewage sludge by thermal behavior, kinetics, thermodynamic parameters and artificial neural
network. Waste Manag. 2019, 85, 131–140. [CrossRef]
46. Ghiba, L.; Drăgoi, E.N.; Curteanu, S. Neural network-based hybrid models developed for free radical polymerization of styrene.
Polym. Eng. Sci. 2021, 61, 716–730. [CrossRef]
47. Liu, Y.-P.; Wu, M.-G.; Qian, J.-X. Evolving Neural Networks Using the Hybrid of Ant Colony Optimization and BP Algorithms.
In Advances in Neural Networks—ISNN 2006; Wang, J., Yi, Z., Zurada, J.M., Lu, B.-L., Yin, H., Eds.; Lecture Notes in Computer
Science; Springer: Berlin/Heidelberg, Germany, 2006; Volume 3971, pp. 714–722; ISBN 978-3-540-34439-1.
48. Liu, Y.; Wu, M.; Qian, J. Predicting coal ash fusion temperature based on its chemical composition using ACO-BP neural network.
Thermochim. Acta 2007, 454, 64–68. [CrossRef]
49. Shao, R.; Martin, E.; Zhang, J.; Morris, A. Confidence bounds for neural network representations. Comput. Chem. Eng. 1997, 21,
S1173–S1178. [CrossRef]
50. Sridhar, D.V.; Seagrave, R.C.; Bartlett, E.B. Process modeling using stacked neural networks. AIChE J. 1996, 42, 2529–2539. [CrossRef]
51. Niaei, A.; Towfighi, J.; Khataee, A.R.; Rostamizadeh, K. The Use of ANN and the Mathematical Model for Prediction of the Main
Product Yields in the Thermal Cracking of Naphtha. Pet. Sci. Technol. 2007, 25, 967–982. [CrossRef]
52. Nielsen, M. Chapter 1: Using Neural Nets to Recognize Handwritten Digits. In Neural Networks and Deep Learning; Determination
Press: San Francisco, CA, USA, 2019.
53. Vapnik, V.N. The Nature of Statistical Learning Theory; Springer: New York, NY, USA, 2000; ISBN 978-1-4419-3160-3.
54. Molga, E. Neural network approach to support modelling of chemical reactors: Problems, resolutions, criteria of application.
Chem. Eng. Process. Process Intensif. 2003, 42, 675–695. [CrossRef]
55. Agnol, L.D.; Ornaghi, H.L., Jr.; Monticeli, F.; Dias, F.T.G.; Bianchi, O. Polyurethanes synthetized with polyols of distinct
molar masses: Use of the artificial neural network for prediction of degree of polymerization. Polym. Eng. Sci. 2021, 61,
1810–1818. [CrossRef]
56. Conesa, J.A.; Caballero, J.; Labarta, J.A. Artificial neural network for modelling thermal decompositions. J. Anal. Appl. Pyrolysis
2004, 71, 343–352. [CrossRef]
57. Ventura, S.; Silva, M.; Perez-Bendito, D.; Hervas, C. Artificial Neural Networks for Estimation of Kinetic Analytical Parameters.
Anal. Chem. 1995, 67, 1521–1525. [CrossRef]
58. Casier, B.; Carniato, S.; Miteva, T.; Capron, N.; Sisourat, N. Using principal component analysis for neural network high-
dimensional potential energy surface. J. Chem. Phys. 2020, 152, 234103. [CrossRef]
59. Ravi Kumar, G.; Nagamani, K.; Anjan Babu, G. A Framework of Dimensionality Reduction Utilizing PCA for Neural Net-work
Prediction. In Advances in Data Science and Management; Borah, S., Emilia Balas, V., Polkowski, Z., Eds.; Lecture Notes on Data
Engineering and Communications Technologies; Springer: Singapore, 2020; Volume 37, pp. 173–180; ISBN 9789811509773.
60. Strange, H.; Zwiggelaar, R. Spectral Dimensionality Reduction. In An Introduction to Distance Geometry applied to Molecular
Geometry; Springer: Berlin/Heidelberg, Germany, 2014; pp. 7–22.
61. Sunphorka, S.; Chalermsinsuwan, B.; Piumsomboon, P. Artificial neural network model for the prediction of kinetic parameters
of biomass pyrolysis from its constituents. Fuel 2017, 193, 142–158. [CrossRef]
62. Zhu, Q.; Jones, J.; Williams, A.; Thomas, K. The predictions of coal/char combustion rate using an artificial neural network
approach. Fuel 1999, 78, 1755–1762. [CrossRef]
63. Bhuyan, N.; Narzari, R.; Baruah, S.M.B.; Kataki, R. Comparative assessment of artificial neural network and response surface
methodology for evaluation of the predictive capability on bio-oil yield of Tithonia diversifolia pyrolysis. Biomass Convers.
Biorefinery 2020. [CrossRef]
64. Dubdub, I.; Al-Yaari, M. Pyrolysis of Low Density Polyethylene: Kinetic Study Using TGA Data and ANN Prediction. Polymers
2020, 12, 891. [CrossRef]
Molecules 2021, 26, 3727 23 of 25

65. Angın, D.; Tiryaki, A.E. Application of response surface methodology and artificial neural network on pyrolysis of safflower seed
press cake. Energy Sources Part A Recover. Util. Environ. Eff. 2016, 38, 1055–1061. [CrossRef]
66. Karacı, A.; Caglar, A.; Aydinli, B.; Pekol, S. The pyrolysis process verification of hydrogen rich gas (H–rG) production by artificial
neural network (ANN). Int. J. Hydrogen Energy 2016, 41, 4570–4578. [CrossRef]
67. Carsky, M.; Kuwornoo, D.K. Neural network modelling of coal pyrolysis. Fuel 2001, 80, 1021–1027. [CrossRef]
68. Al-Yaari, M.; Dubdub, I. Application of Artificial Neural Networks to Predict the Catalytic Pyrolysis of HDPE Using Non-
Isothermal TGA Data. Polymers 2020, 12, 1813. [CrossRef]
69. Arumugasamy, S.K.; Selvarajoo, A. Feedforward Neural Network Modeling of Biomass Pyrolysis Process for Biochar Produc-tion.
Chem. Eng. Trans. 2015, 45, 1681–1686. [CrossRef]
70. Li, X.H.; Fan, Y.S.; Cai, Y.X.; Zhao, W.D.; Yin, H.Y. Optimization of Biomass Vacuum Pyrolysis Process Based on GRNN. Appl.
Mech. Mater. 2013, 411–414, 3016–3022. [CrossRef]
71. Bi, H.; Wang, C.; Lin, Q.; Jiang, X.; Jiang, C.; Bao, L. Combustion behavior, kinetics, gas emission characteristics and artificial
neural network modeling of coal gangue and biomass via TG-FTIR. Energy 2020, 213, 118790. [CrossRef]
72. Sathiya Prabhakaran, S.P.; Swaminathan, G.; Joshi, V.V. Thermogravimetric analysis of hazardous waste: Pet-coke, by kinetic
models and Artificial neural network modeling. Fuel 2021, 287, 119470. [CrossRef]
73. Bi, H.; Wang, C.; Jiang, X.; Jiang, C.; Bao, L.; Lin, Q. Prediction of mass loss for sewage sludge-peanut shell blends in thermogravi-
metric experiments using artificial neural networks. Energy Sources Part A Recover. Util. Environ. Eff. 2020, 1–14. [CrossRef]
74. Chen, J.; Xie, C.; Liu, J.; He, Y.; Xie, W.; Zhang, X.; Chang, K.; Kuo, J.; Sun, J.; Zheng, L.; et al. Co-combustion of sewage sludge
and coffee grounds under increased O2 /CO2 atmospheres: Thermodynamic characteristics, kinetics and artificial neural network
modeling. Bioresour. Technol. 2018, 250, 230–238. [CrossRef]
75. Monticeli, F.M.; Neves, R.M.; Júnior, H.L.O. Using an artificial neural network (ANN) for prediction of thermal degradation from
kinetics parameters of vegetable fibers. Cellulose 2021, 28, 1961–1971. [CrossRef]
76. Gu, C.; Wang, X.; Song, Q.; Li, H.; Qiao, Y. Prediction of gas-liquid-solid product distribution after solid waste pyrolysis process
based on artificial neural network model. Int. J. Energy Res. 2021, 6707. [CrossRef]
77. Saleem, M.; Ali, I. Machine Learning Based Prediction of Pyrolytic Conversion for Red Sea Seaweed. In Proceedings of the 7th In-
ternational Conference on Biological, Chemical & Environmental Sciences (BCES-2017), Budapest, Hungary, 6–7 September 2017.
78. Abdurakipov, S.S.; Butakov, E.B.; Burdukov, A.P.; Kuznetsov, A.V.; Chernova, G.V. Using an Artificial Neural Network to Simulate
the Complete Burnout of Mechanoactivated Coal. Combust. Explos. Shock. Waves 2019, 55, 697–701. [CrossRef]
79. Cao, H.; Xin, Y.; Yuan, Q. Prediction of biochar yield from cattle manure pyrolysis via least squares support vector machine
intelligent approach. Bioresour. Technol. 2016, 202, 158–164. [CrossRef] [PubMed]
80. Çepelioğullar, Ö.; Mutlu, I.; Yaman, S.; Haykiri-Acma, H. Activation energy prediction of biomass wastes based on different
neural network topologies. Fuel 2018, 220, 535–545. [CrossRef]
81. Aydinli, B.; Caglar, A.; Pekol, S.; Karaci, A. The prediction of potential energy and matter production from biomass pyrolysis
with artificial neural network. Energy Explor. Exploit. 2017, 35, 698–712. [CrossRef]
82. Burgaz, E.; Yazici, M.; Kapusuz, M.; Alisir, S.H.; Ozcan, H. Prediction of thermal stability, crystallinity and thermomechan-
ical properties of poly(ethylene oxide)/clay nanocomposites with artificial neural networks. Thermochim. Acta 2014, 575,
159–166. [CrossRef]
83. Himmelblau, D.M. Applications of artificial neural networks in chemical engineering. Korean J. Chem. Eng. 2000, 17,
373–392. [CrossRef]
84. Molga, E.J.; van Woezik, B.A.A.; Westerterp, K.R. Neural networks for modelling of chemical reaction systems with complex
kinetics: Oxidation of 2-octanol with nitric acid. Chem. Eng. Process. Process Intensif. 2000, 39, 323–334. [CrossRef]
85. Von Stosch, M.; Oliveira, R.; Peres, J.; de Azevedo, S.F. Hybrid semi-parametric modeling in process systems engineering: Past,
present and future. Comput. Chem. Eng. 2014, 60, 86–101. [CrossRef]
86. Bing, G. Modelling coal gasification with a hybrid neural network. Fuel 1997, 76, 1159–1164. [CrossRef]
87. Roduit, B.; Borgeat, C.; Berger, B.; Folly, P.; Alonso, B.; Aebischer, J.N. The prediction of thermal stability of self-reactive chemicals:
From milligrams to tons. J. Therm. Anal. Calorim. 2005, 80, 91–102. [CrossRef]
88. Agarwal, M. Combining neural and conventional paradigms for modelling, prediction and control. Int. J. Syst. Sci. 1997, 28,
65–81. [CrossRef]
89. Rojek, B.; Suchacz, B.; Wesolowski, M. Artificial neural networks as a supporting tool for compatibility study based on thermo-
gravimetric data. Thermochim. Acta 2018, 659, 222–231. [CrossRef]
90. Kohonen, T. Self-Organizing Maps; Springer Series in Information Sciences; Springer: Berlin/Heidelberg, Germany, 2001;
Volume 30; ISBN 978-3-540-67921-9.
91. Strzelczak, A. The application of artificial neural networks (ANN) for the denaturation of meat proteins—The kinetic analysis
method. Acta Sci. Pol. Technol. Aliment. 2015, 18, 87–96. [CrossRef]
92. Charde, S.J. Degradation Kinetics of Polycarbonate Composites: Kinetic Parameters and Artificial Neural Network. Chem.
Biochem. Eng. Q. 2018, 32, 151–165. [CrossRef]
93. Ferreira, B.D.L.; Araujo, B.C.R.; Sebastião, R.C.O.; Yoshida, M.I.; Mussel, W.N.; Fialho, S.; Barbosa, J. Kinetic study of anti-HIV
drugs by thermal decomposition analysis: A multilayer artificial neural network propose. J. Therm. Anal. Calorim. 2017, 127,
577–585. [CrossRef]
Molecules 2021, 26, 3727 24 of 25

94. Huang, Y.W.; Chen, M.Q.; Li, Q.H. Artificial neural network model for the evaluation of chemical kinetics in thermally induced
solid-state reaction. J. Therm. Anal. Calorim. 2019, 138, 451–460. [CrossRef]
95. Kuang, D.; Xu, B. Predicting kinetic triplets using a 1d convolutional neural network. Thermochim. Acta 2018, 669, 8–15. [CrossRef]
96. Aghbashlo, M.; Almasi, F.; Jafari, A.; Nadian, M.H.; Soltanian, S.; Lam, S.S.; Tabatabaei, M. Describing biomass pyrolysis kinetics
using a generic hybrid intelligent model: A critical stage in sustainable waste-oriented biorefineries. Renew. Energy 2021, 170,
81–91. [CrossRef]
97. Aghbashlo, M.; Tabatabaei, M.; Nadian, M.H.; Davoodnia, V.; Soltanian, S. Prognostication of lignocellulosic biomass pyrolysis
behavior using ANFIS model tuned by PSO algorithm. Fuel 2019, 253, 189–198. [CrossRef]
98. Jang, J.-S. ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans. Syst. Man Cybern. 1993, 23, 665–685. [CrossRef]
99. Rouquerol, J. Sample-Controlled Thermal Analysis. In Reference Module in Chemistry, Molecular Sciences and Chemical Engineering;
Elsevier: Amsterdam, The Netherlands, 2018; p. 9780124095472144000.
100. Hu, X.; Lin, Z.; Yang, K.; Deng, Z. Kinetic Analysis of One-Step Solid-State Reaction for Li4 Ti5 O12 /C. J. Phys. Chem. A 2011, 115,
13413–13419. [CrossRef] [PubMed]
101. De Freitas-Marques, M.B.; Araujo, B.C.R.; Da Silva, P.H.R.; Fernandes, C.; Mussel, W.D.N.; Sebastião, R.D.C.D.O.; Yoshida, M.I.
Multilayer perceptron network and Vyazovkin method applied to the non-isothermal kinetic study of the interaction between
lumefantrine and molecularly imprinted polymer. J. Therm. Anal. Calorim. 2020. [CrossRef]
102. Vyazovkin, S.; Burnham, A.K.; Favergeon, L.; Koga, N.; Moukhina, E.; Pérez-Maqueda, L.A.; Sbirrazzuoli, N. ICTAC Kinetics
Committee recommendations for analysis of multi-step kinetics. Thermochim. Acta 2020, 689, 178597. [CrossRef]
103. Ferreira, B.D.L.; Araújo, N.R.S.; Ligório, R.F.; Pujatti, F.J.P.; Mussel, W.N.; Yoshida, M.I.; Sebastião, R.C.O. Kinetic thermal
decomposition studies of thalidomide under non-isothermal and isothermal conditions. J. Therm. Anal. Calorim. 2018, 134,
773–782. [CrossRef]
104. Ferreira, B.D.D.L.; Araújo, N.R.; Ligório, R.F.; Pujatti, F.J.; Yoshida, M.I.; Sebastião, R.C. Comparative kinetic study of automotive
polyurethane degradation in non-isothermal and isothermal conditions using artificial neural network. Thermochim. Acta 2018,
666, 116–123. [CrossRef]
105. Ferreira, L.D.L.; Medeiros, F.S.; Araujo, B.C.; Gomes, M.S.; Rocco, M.L.M.; Sebastião, R.C.; Calado, H.D. Kinetic study of MWCNT
and MWCNT@P3HT hybrid thermal decomposition under isothermal and non-isothermal conditions using the artificial neural
network and isoconversional methods. Thermochim. Acta 2019, 676, 145–154. [CrossRef]
106. Vieira, S.S.; Paz, G.M.; Araujo, B.C.; Lago, R.M.; Sebastião, R.C. Use of neural network to analyze the kinetics of CO2 absorption
in Li4 SiO4 /MgO composites from TG experimental data. Thermochim. Acta 2020, 689, 178628. [CrossRef]
107. Marques, M.B.D.F.; De Araujo, B.C.R.; Sebastião, R.D.C.D.O.; Mussel, W.D.N.; Yoshida, M.I. Kinetics study and Hirshfeld surface
analysis for atorvastatin calcium trihydrate and furosemide system. Thermochim. Acta 2019, 682, 178408. [CrossRef]
108. Khawam, A.; Flanagan, D.R. Solid-State Kinetic Models: Basics and Mathematical Fundamentals. J. Phys. Chem. B 2006, 110,
17315–17328. [CrossRef] [PubMed]
109. Vyazovkin, S.; Wight, C.A. Model-free and model-fitting approaches to kinetic analysis of isothermal and nonisothermal data.
Thermochim. Acta 1999, 340–341, 53–68. [CrossRef]
110. Rahman, M.S.; Rashid, M.; Hussain, M. Thermal conductivity prediction of foods by Neural Network and Fuzzy (ANFIS)
modeling techniques. Food Bioprod. Process. 2012, 90, 333–340. [CrossRef]
111. Ramezanizadeh, M.; Nazari, M.A.; Ahmadi, M.H.; Lorenzini, G.; Pop, I. A review on the applications of intelligence methods in
predicting thermal conductivity of nanofluids. J. Therm. Anal. Calorim. 2019, 138, 827–843. [CrossRef]
112. Lazzús, J.A. Prediction of solid vapor pressures for organic and inorganic compounds using a neural network. Thermochim. Acta
2009, 489, 53–62. [CrossRef]
113. Asante-Okyere, S.; Xu, Q.; Mensah, R.A.; Jin, C.; Ziggah, Y.Y. Generalized regression and feed forward back propagation
neural networks in modelling flammability characteristics of polymethyl methacrylate (PMMA). Thermochim. Acta 2018, 667,
79–92. [CrossRef]
114. Shahbaz, K.; Baroutian, S.; Mjalli, F.S.; Hashim, M.A.; AlNashef, I.M. Densities of ammonium and phosphonium based deep
eutectic solvents: Prediction using artificial intelligence and group contribution techniques. Thermochim. Acta 2012, 527,
59–66. [CrossRef]
115. Lisa, G.; Wilson, D.A.; Curteanu, S.; Lisa, C.; Piuleac, C.-G.; Bulacovschi, V. Ferrocene derivatives thermostability prediction
using neural networks and genetic algorithms. Thermochim. Acta 2011, 521, 26–36. [CrossRef]
116. Kopal, I.; Harničárová, M.; Valíček, J.; Kušnerová, M. Modeling the Temperature Dependence of Dynamic Mechanical Prop-
erties and Visco-Elastic Behavior of Thermoplastic Polyurethane Using Artificial Neural Network. Polymers 2017, 9, 519.
[CrossRef] [PubMed]
117. Gao, F.; Wang, F.; Li, M. Neural network-based optimal iterative controller for nonlinear processes. Can. J. Chem. Eng. 2000, 78,
363–370. [CrossRef]
118. Chen, L.; Ge, B.; de Almeida, A.T. Self-tuning PID Temperature Controller Based on Flexible Neural Network. In Advances in
Neural Networks—ISNN 2007; Liu, D., Fei, S., Hou, Z.-G., Zhang, H., Sun, C., Eds.; Lecture Notes in Computer Science; Springer:
Berlin/Heidelberg, Germany, 2007; Volume 4491, pp. 138–147. ISBN 978-3-540-72382-0.
119. Khalid, M.; Omatu, S. A neural network controller for a temperature control system. IEEE Control Syst. 1992, 12, 58–64. [CrossRef]
120. Townsend, D.I.; Tou, J.C. Thermal hazard evaluation by an accelerating rate calorimeter. Thermochim. Acta 1980, 37, 1–30. [CrossRef]
Molecules 2021, 26, 3727 25 of 25

121. Voga, G.P.; Belchior, J.C. An approach for interpreting thermogravimetric profiles using artificial intelligence. Thermochim. Acta
2007, 452, 140–148. [CrossRef]
122. Voga, G.P.; Coelho, M.G.; De Lima, G.M.; Belchior, J.C. Experimental and Theoretical Studies of the Thermal Behavior of Titanium
Dioxide−SnO2 Based Composites. J. Phys. Chem. A 2011, 115, 2719–2726. [CrossRef]
123. Lerkkasemsan, N.; Achenie, L.E. Pyrolysis of biomass—Fuzzy modeling. Renew. Energy 2014, 66, 747–758. [CrossRef]
124. Fazilat, H.; Akhlaghi, S.; Shiri, M.E.; Sharif, A. Predicting thermal degradation kinetics of nylon6/feather keratin blends using
artificial intelligence techniques. Polymer 2012, 53, 2255–2264. [CrossRef]
125. Pathy, A.; Meher, S.; Balasubramanian, P. Predicting algal biochar yield using eXtreme Gradient Boosting (XGB) algorithm of
machine learning methods. Algal Res. 2020, 50, 102006. [CrossRef]